الذكاء الاصطناعي التوليدي: تطور التكنولوجيا، التأثير الاجتماعي المتزايد، والفرص لأبحاث نظم المعلومات Generative Artificial Intelligence: Evolving Technology, Growing Societal Impact, and Opportunities for Information Systems Research

المجلة: Information Systems Frontiers
DOI: https://doi.org/10.1007/s10796-025-10581-7
تاريخ النشر: 2025-02-25

الذكاء الاصطناعي التوليدي: تطور التكنولوجيا، التأثير الاجتماعي المتزايد، والفرص لأبحاث نظم المعلومات

فيدا سي. ستوري وي ثو يوي ج. ليون تشاو (د) ⋅ رومان لوكيانينكو (د)

تم القبول: 7 يناير 2025
© المؤلفون 2025

الملخص

أدت التطورات المستمرة والانفجارية في الذكاء الاصطناعي التوليدي (GenAI)، المبني على نماذج اللغة الكبيرة والخوارزميات ذات الصلة، إلى الكثير من الحماس والتكهنات حول التأثير المحتمل لهذه التكنولوجيا الجديدة. تشمل الادعاءات أن الذكاء الاصطناعي (AI) على وشك إحداث ثورة في الأعمال والمجتمع وتغيير الحياة الشخصية بشكل جذري. ومع ذلك، ليس من الواضح كيف أن هذه التكنولوجيا، بميزاتها المميزة بشكل كبير عن تقنيات الذكاء الاصطناعي السابقة، لديها القدرة التحويلية أو كيف يجب أن يتفاعل الباحثون في نظم المعلومات معها. في هذه الورقة، نعتبر الاتجاهات المتطورة والناشئة للذكاء الاصطناعي من أجل فحص تأثيراته الحالية والتنبؤ بتأثيراته المستقبلية. العديد من الأوراق الموجودة حول GenAI إما أنها تقنية للغاية بالنسبة لمعظم الباحثين في نظم المعلومات أو تفتقر إلى العمق اللازم لتقدير التأثيرات المحتملة لـ GenAI. لذلك، نحاول سد الفجوة بين المجتمعات التقنية والتنظيمية لـ GenAI من منظور اجتماعي تقني موجه نحو النظام. على وجه التحديد، نستكشف الميزات الفريدة لـ GenAI، التي تتجذر في التغيير المستمر من الرمزية إلى الاتصال، والخصائص النظامية العميقة والفطرية لنظم الإنسان-الذكاء الاصطناعي. نتتبع تطور الذكاء الاصطناعي الذي سبق مستوى التبني والتكيف والاستخدام الموجود اليوم، من أجل اقتراح أبحاث مستقبلية حول التأثيرات المختلفة لـ GenAI في كل من الأعمال والمجتمع في سياق أبحاث نظم المعلومات. تهدف جهودنا إلى المساهمة في إنشاء جدول أبحاث منظم جيدًا في مجتمع نظم المعلومات لدعم استراتيجيات وعمليات مبتكرة ممكنة بفضل هذه الموجة الجديدة من الذكاء الاصطناعي.

الكلمات الرئيسية: الذكاء الاصطناعي (AI) • الذكاء الاصطناعي التوليدي (GenAI) • نماذج اللغة الكبيرة (LLM) • الاتصال • التأثير الاجتماعي • إطار عمل للذكاء الاصطناعي التوليدي كنظام اجتماعي تقني • نظم المعلومات

1 المقدمة

يعتبر الذكاء الاصطناعي تكنولوجيا تحويلية بشكل هائل، تؤثر على المنظمات والأفراد بطرق يصعب فهمها بالكامل دون الاستفادة من نظرة تاريخية طويلة الأمد. يشير الذكاء الاصطناعي (AI) إلى قدرة الحواسيب على أداء المهام التي كانت تاريخيًا تتطلب الإدراك البشري وغيرها من القدرات الفكرية
مثل الإدراك، التجريد، الاستنتاج، التعلم، واتخاذ القرار (راسل ونورفيغ، 2016). يُعتبر الذكاء الاصطناعي بشكل متزايد قمة العلوم والهندسة البشرية (راسل ونورفيغ، 2016). التقديرات الصناعية تشير إلى أن القيمة الاقتصادية العالمية للذكاء الاصطناعي ستصل إلى 15 تريليون دولار بحلول عام 2030 (PwC، 2024). لقد تم الاعتراف منذ فترة طويلة بأهمية البحث في الذكاء الاصطناعي. وقد تم اقتراح حتى أن الدولة التي تتصدر العالم في الذكاء الاصطناعي قد تظهر كقوة رائدة في العالم (سافاج، 2020).
تقدم الذكاء الاصطناعي اتبع سلسلة من التحولات النموذجية الهامة، تعكس الاتجاهات الأوسع في الابتكار التكنولوجي وانتشاره (بيريز، 2010). قبل العقد الأول من القرن الحادي والعشرين، شهد مجال الذكاء الاصطناعي “شتاء الذكاء الاصطناعي” بسبب الركود في التقدم العملي، حيث كانت أبحاث الذكاء الاصطناعي تهيمن عليها بشكل أساسي طرق التعلم المستندة إلى المنطق والنماذج، والتجريد، وطرق الاستنتاج (كريفير، 1993). منذ العقد الأول من القرن الحادي والعشرين، أدت وفرة البيانات الكبيرة والمتنوعة لتدريب، والتوسع المستمر في قوة الحوسبة، والتقدم في تصميم خوارزميات الذكاء الاصطناعي، إلى تحويل التركيز تقريبًا
حصريًا نحو التعلم الآلي (ML) والذكاء الاصطناعي المستند إلى البيانات (سيرف، 2019)، حيث يتكون التعلم الآلي من طرق وخوارزميات تُستخدم لاستنتاج المعلومات من البيانات (غودفيلو وآخرون، 2020). مع اعتمادها الواسع، بدأ التعلم الآلي في تحويل المنظمات وحتى الصناعات بأكملها بسرعة.
كان الباحثون في نظم المعلومات (IS) دائمًا في طليعة البحث حول اعتماد وتقدم التقنيات الجديدة من أجل ابتكار ممارسات الأعمال والاجتماع. تشمل بعض الأمثلة على الأجيال السابقة من تقنيات المعلومات التي تم التحقيق فيها بعناية من قبل نظم المعلومات معالجة البيانات التنظيمية، واتخاذ القرارات المستندة إلى البيانات والتحليلات، وتخطيط موارد المؤسسات، والتجارة الإلكترونية. نحن الآن نشهد ثورة أخرى في الذكاء الاصطناعي ناتجة عن التطورات الأخيرة في الشبكات العصبية العميقة (شمدهوبر، 2015) ومعالجة اللغة الطبيعية (روتمان، 2023). يشير الذكاء الاصطناعي التوليدي (GenAI) إلى أنظمة الذكاء الاصطناعي القادرة على إنتاج محتوى مثل النصوص، والصور، والموسيقى، وبرامج الكود، وغيرها من المخرجات المعقدة والإبداعية. GenAI ليس جديدًا. شملت GenAI السابقة، على سبيل المثال، الشبكات التنافسية التوليدية (غودفيلو وآخرون، 2014). ومع ذلك، فإن التقدم الأخير في GenAI مدفوع بالأداء المثير للإعجاب لنماذج اللغة الكبيرة (LLMs). LLMs هي نماذج حسابية لديها القدرة على فهم وتوليد اللغة البشرية من خلال القدرة التحويلية على “توقع احتمالية تسلسل الكلمات أو توليد نص جديد بناءً على مدخلات معينة” (تشانغ وآخرون، 2024، ص.2). بينما تحاول LLMs تقديم قدرات شبيهة بالبشر، فإنها تواجه أيضًا العديد من التحديات بسبب الكلمات النادرة أو غير المرئية، والتكيف المفرط، والقواعد اللغوية المعقدة. وبالتالي، تستمر ظهور هياكل جديدة لـ LLM وطرق التدريب للتغلب على هذه التحديات وغيرها (تشانغ وآخرون، 2024).
نتيجة التقدم في GenAI هي الكثير من الاهتمام والتكهنات بشأن دور الذكاء الاصطناعي للاستخدام اليومي، الذي تم تحفيزه جزئيًا من خلال الشعبية المتزايدة لأدوات مثل ChatGPT وDall-E (open.ai), MidJourney (midjourney.com), Google Bard (bard.google.com), وCoPilot AI (copilotai.com). ChatGPT، هو مثال على GenAI يمكنه معالجة كميات كبيرة من البيانات لإنشاء محتوى جديد (مثل: صورة، مقال، كود برنامج، كلمات أغاني). من الناحية الأساسية، تجعل LLMs من الممكن لأنظمة الذكاء الاصطناعي تقييم والتفاعل مع السياقات البشرية بطرق قابلة للتواصل ومعقولة (بيندر وآخرون، 2021)، مما يسمح لهذه الأنظمة بأداء مهام مركزية للمعرفة لم تكن قادرة على إنجازها سابقًا. تعتبر هذه القدرات جديدة، مع إمكانية دعم حالات استخدام تجارية لا حصر لها. تتراوح المراجعات
والتحليلات لأدوات GenAI، مثل ChatGPT، من “عبور عتبة”، حيث يمكن استخدام هذه الأداة العامة لمجموعة واسعة من الأنشطة (مثل: طرق جديدة وفعالة لأداء العمل) إلى الانتحال على نطاق واسع والمعلومات المضللة (سافاج، 2023؛ سوسارلا وآخرون، 2023).
أدى ظهور GenAI إلى العديد من الأسئلة غير المجابة حول كيفية استجابة المجتمع بشكل عام لهذه التكنولوجيا الجديدة، لا سيما فيما يتعلق بأدوارها وآثارها. منذ ظهور ChatGPT وأنظمة GenAI الأخرى، أدى التأثير الاجتماعي لـ GenAI إلى طرح أسئلة مثل: كيف سيحول GenAI، كنظم تقنية، الأنشطة التجارية؟ ما التحديات التي يجب تحديدها وحلها لضمان الاستخدام المنتج والأخلاقي والآمن والمسؤول لهذه التكنولوجيا؟ هذه الفجوة في فهم قيمة ومخاطر GenAI تؤدي إلى سؤال بحثنا الرئيسي: ما المنظور الفريد والقيم الذي يمكن أن تقدمه أبحاث نظم المعلومات فيما يتعلق بتكنولوجيا GenAI وتأثيراتها على الأفراد والمنظمات؟ نظرًا لأن البحث في نظم المعلومات كان عمومًا بارعًا في دراسة الترابط بين الجوانب الاجتماعية والتقنية للأنظمة (الطبيعة الاجتماعية التقنية للتقدم في التكنولوجيا)، يجب أن يكون أيضًا ذا صلة وقادرًا على دراسة والمساهمة في فهم قيمة ومخاطر GenAI.
تقدم هذه الورقة مفهوم “النظام الاجتماعي التقني التوليدي” لاستكشاف ما يجعل الذكاء الاصطناعي التوليدي (GenAI) تحويليًا ولدعم انتشاره المستمر في المنظمات والمجتمع. نطور إطارًا نظريًا يعتمد على نظرية الأنظمة (بما في ذلك عدسة اجتماعية تقنية) ومدعومًا بأسس نظرية أخرى (مثل نظرية اللغة). يتم استخدام الأساس النظامي لوضع مكونات GenAI وسلوكها ويساعد في تطوير أجندة بحثية لدراسة نظم المعلومات تستند إلى القوى والقدرات الموجودة في تخصصنا. كما يمكّننا استخدام نظرية الأنظمة من دراسة كيفية استفادة المنظمات من تكنولوجيا GenAI، وكيف يمكن دمج GenAI في العمليات الحالية بشكل مسؤول ومربح، وكيفية ربط هذه التكنولوجيا الجديدة بفعالية مع الموارد التنظيمية الحالية. بهذه الطريقة، يجب أن تركز دراسة نظم المعلومات على استكشاف طبيعة وآثار اعتبار GenAI كعنصر من عناصر الأنظمة الاجتماعية التقنية الأوسع، بدلاً من التركيز فقط على الطبيعة التقنية لـ GenAI نفسها.
تقدم هذه الدراسة عدة مساهمات. أولاً، نقترح إطارًا نظريًا لفهم GenAI (إطار الذكاء الاصطناعي التوليدي كنظام اجتماعي تقني)، من منظور اجتماعي تقني، متسق مع الأبحاث الحالية في نظم المعلومات. ثانيًا، نظهر أن GenAI يتقدم إلى نقطة حيث سيكون للأدوات القدرة على توليد نتائج غير متوقعة، مما يتطلب شروط حدودية وطرق جديدة للنظر في المبادرات البحثية وتحليلها. ثالثًا، نقترح مفهوم النظام الاجتماعي التقني التوليدي. من خلال تصور
طبيعة GenAI من منظور الأنظمة، يمكننا تطوير فهم دقيق لـ GenAI من خلال الاستفادة من المفاهيم والنظريات الموجودة في أبحاث نظم المعلومات. رابعًا، نحدد مواضيع مثمرة للبحث المستقبلي.
تتقدم هذه الورقة على النحو التالي. يستعرض القسم 2 التقدم في الذكاء الاصطناعي على مدى العقود الثلاثة الماضية من أجل وضع فهمنا لتقدم GenAI. يقدم القسم 3 إطارًا نظريًا لفهم GenAI. يقترح القسم 4 مواضيع بحثية محتملة بناءً على الإطار، ويختتم القسم 5 الورقة.
هناك مجموعة متزايدة من الأبحاث ضمن مجالات نظم المعلومات (IS) والأعمال التي تسعى لفهم وتحليل فوائد وآثار الذكاء الاصطناعي التوليدي (GenAI) من وجهات نظر مختلفة. اعتمد دويفيدي وآخرون (2023) منظورًا متعدد التخصصات لتفصيل الآثار المتعددة الأوجه لـ GenAI على المجتمع. على الرغم من وعوده، يشير فوي-هون ناه (فويهون ناه وآخرون، 2023) إلى التحديات الأخلاقية والتكنولوجية والتنظيمية المرتبطة بـ GenAI. ضمن مجال نظم المعلومات، استكشفت الأبحاث تأثير GenAI على التعليم (كاجتازي وآخرون، 2023). يناقش جاريفابا وكلاين (2024) إمكانية GenAI للمساعدة في بناء النظرية
في مجال نظم المعلومات. يقترح سوسارلا وآخرون (2023) إمكانية الاستفادة من GenAI في قدرات مختلفة لإجراء العمل الأكاديمي. في سياق أوسع، يجادل سابهر وال وغروفر (2024) بأن التأثير الاجتماعي للذكاء الاصطناعي يعتمد على تطويره وتنفيذه. يؤكدون على أهمية النظر في مدى استبدال الذكاء الاصطناعي للبشر مقابل دعمه لهم، ودمج الحقائق المادية والرقمية، واحترام القيود البشرية. يحدد أوليري (2022) القضايا الناشئة المرتبطة بالنماذج اللغوية الكبيرة (LLMs) بشكل عام. يقدم العلاوي وآخرون (2024) اقتراحات حول كيفية تركيز مجال نظم المعلومات على دور GenAI من منظور إدارة المعرفة، مع تحديد فرص البحث ذات الصلة. نواصل هذه الجهود من خلال تقديم تحليل شامل وعام لـ GenAI كتكنولوجيا ونقترح أجندة مثمرة لدراسة نظم المعلومات.

2.1 الجهود السابقة في البحث في الذكاء الاصطناعي ونظم المعلومات

لتقييم تأثير GenAI، بالإضافة إلى المبادرات وراء الذكاء الاصطناعي وقدراته، من الضروري فهم تطور الذكاء الاصطناعي والتقدم التكنولوجي الذي أدى إلى انتعاشه. يوضح هذا الفهم الإمكانيات والقيود المتعلقة بالذكاء الاصطناعي. توضح الشكل 1 تقدم الذكاء الاصطناعي والمساهمات الكبيرة التي قدمها الباحثون في نظم المعلومات مع تقدم الذكاء الاصطناعي
الشكل 1 تقدم الذكاء الاصطناعي والأبحاث ذات الصلة في نظم المعلومات
الجدول 1 ملخص للإنجازات المبكرة في الذكاء الاصطناعي
السنة أداة الذكاء الاصطناعي أهمية الذكاء الاصطناعي المرجع
1952 لعبة الداما أثبتت أن الحواسيب يمكن أن تتعلم اللعب بمستوى عالٍ يكفي لتحدي لاعب بشري هاوٍ صموئيل (1960)
1955 المفكر المنطقي أثبت 38 نظرية من مبادئ الرياضيات؛ قدم مفاهيم حاسمة في الذكاء الاصطناعي (مثل، الاستدلال، معالجة القوائم، التفكير كبحث) نيويل وآخرون (1962)
1957 المدرك ولادة الاتصال؛ أساس الشبكات العصبية (NN)، التعلم العميق روزنبلات (1961)
1961 MENACE أول برنامج قادر على تعلم اللعب بشكل مثالي في لعبة إكس-أو ميشي (1963)
1965 ELIZA نظام معالجة اللغة الطبيعية الذي قلّد الطبيب من خلال الرد على أسئلة مشابهة للمعالج النفسي قبل أن تصبح المحادثة غير منطقية فايتنباوم (1966)
1969 شاكِي الروبوت أول روبوت متنقل عام قادر على التفكير، دمج البحث في الروبوتات مع رؤية الكمبيوتر ومعالجة اللغة الطبيعية بيرترام (1972)
1969 الكتاب “المدركات” سلط الضوء على الحدود غير المعترف بها لبنية المدرك ذو الطبقتين؛ تحول أساسي في أبحاث الذكاء الاصطناعي نحو الرمزية، متجاهلاً الاتصال مينسكي وبابرت (1969)
من أعماله المبكرة إلى GenAI. الجزء أعلاه الخط يظهر المراحل الرئيسية للابتكارات في الذكاء الاصطناعي، وأسفل الخط، المجالات ذات الصلة من التقدم والجهود ذات الصلة في أبحاث نظم المعلومات السابقة. تشير الشكل 1 إلى أن أبحاث نظم المعلومات كانت دائمًا تحاول مواكبة الأبحاث في الذكاء الاصطناعي وتوفر الأساس لهذه الدراسة حول GenAI وتطبيقاته من منظور نظم المعلومات.

2.2 التطور من الذكاء الاصطناعي العام إلى الذكاء الاصطناعي التوليدي

بينما يعد GenAI تطورًا حديثًا، فإن جذوره تعود إلى الأسس الأساسية للذكاء الاصطناعي. كانت الأبحاث الأولية في الذكاء الاصطناعي تهدف إلى إنشاء آلات قادرة على الفهم والتفكير مثل البشر، مع اختبار تورينغ الشهير الذي اقترحه آلان تورينغ في عام 1950 (تورينغ، 1950). في عام 1957، طور هربرت سيمون، جون شو، وآلن نيويل حلاً عامًا للمشكلات، بهدف محاكاة حل المشكلات البشرية. في عام 1959، نشر جون مكارثي ورقة بعنوان: “برامج ذات حس مشترك” (مكارثي، 1959). أصبح تطوير الأنظمة الشاملة التي تسعى لمطابقة الإدراك البشري عبر مجموعة واسعة من المهام يُعرف لاحقًا بالذكاء الاصطناعي العام (AGI) أو الذكاء الاصطناعي العام.
أثبتت أجندة الذكاء الاصطناعي العام أنها أقل نجاحًا مما كان متوقعًا، على الرغم من مستوى الاهتمام والاستثمار العالي (راسل ونورفيغ، 2016). بدلاً من ذلك، أصبحت الأنظمة الخبيرة، التي كانت محاولات قائمة على القواعد لالتقاط وتمثيل معرفة الخبراء البشريين بحيث يمكن إعادة استخدامها، النموذج السائد للذكاء الاصطناعي. كان تطوير الذكاء الاصطناعي العام صعبًا جدًا بالنظر إلى حالة الحوسبة في الثمانينيات. أصبحت الأنظمة الخبيرة القائمة على القواعد والمخصصة للقطاعات هي الخيار لمعظم جهود الذكاء الاصطناعي (سيغيف وزهاو، 1994). مع النماذج اللغوية الكبيرة (LLMs)، أصبحت إمكانية الذكاء الاصطناعي العام أكثر قربًا. على سبيل المثال، تشير شركة OpenAI بشكل خاص إلى أن
منتجها، ChatGPT، يمكنه الإجابة على الأسئلة في أي مجال (أغراوال، 2023). وهذا يشير إلى أن النماذج اللغوية الكبيرة يمكن تدريبها على بيانات ضخمة تشمل العديد من المجالات.
لتقدير تطور الذكاء الاصطناعي من بدايته إلى GenAI الذي يمتد على 70 عامًا من التطور منذ الخمسينيات (ديليبتريف وآخرون، 2020)، نRecognize ثلاثة فترات تطوير رئيسية: أسس الذكاء الاصطناعي (1950-1970)؛ الذكاء الاصطناعي الرمزي (1970-1990)؛ والتعلم الآلي والتعلم العميق (1990-2020).
أسس الذكاء الاصطناعي (الخمسينات – السبعينات). في عام 1950، نشر آلان تورينغ الورقة البحثية الرائدة “الآلات الحاسوبية والذكاء” (تورينغ، 1950) (تورينغ، 2012)، متسائلاً السؤال الأساسي “هل يمكن للآلات أن تفكر؟” اقترح تورينغ لعبة تقليد، عُرفت فيما بعد باسم اختبار تورينغ، حيث إذا كانت الآلة قادرة على إجراء محادثة لا يمكن تمييزها عن محادثة مع إنسان، فمن المعقول أن نستنتج أن الآلة ذكية. أصبح اختبار تورينغ أول تجربة تحاول قياس ذكاء الآلة. أشعل مؤتمر دارتموث في عام 1956 بداية الذكاء الاصطناعي عندما صاغ مكارثي مصطلح “الذكاء الاصطناعي”، مما أدى إلى ظهور هذا المجال العلمي الجديد.
كانت المهمة الأساسية للمجال الجديد للذكاء الاصطناعي هي دراسة “كل جانب من أي ميزة أخرى من التعلم أو الذكاء [لكي] يتم وصفه بدقة حتى تتمكن الآلة من محاكاته” (راسل ونورفيغ، 2016). منذ ذلك الحين، حل مجتمع أبحاث الذكاء الاصطناعي مشاكل مثل مشاكل تطبيقات الجبر، وترجمة اللغات، وإثبات النظريات الهندسية. تلخص الجدول 1 الإنجازات المهمة في الذكاء الاصطناعي خلال تلك الفترة، استنادًا إلى دليبتريف وآخرون (2020).
كما هو موضح في الجدول 1، بدأت المكونات الأساسية لنماذج اللغة الكبيرة، وهي الشبكات العصبية، والتعلم العميق، ومعالجة اللغة الوطنية، في مشاريع تاريخية، مثل بيرسيبترون (1957)، وإليزا (1961)، وشاكي الروبوت (1969)، في الفترة التأسيسية للذكاء الاصطناعي. أي أنه استغرق الأمر حوالي ستين عامًا من إنشاء أسس الذكاء الاصطناعي في المختبر إلى نماذج اللغة الكبيرة المتاحة للجمهور العام.
الذكاء الاصطناعي الرمزي (السبعينات – التسعينات). ركز عصر الذكاء الاصطناعي الرمزي على تطوير أنظمة تعتمد على قواعد تم تنسيقها بواسطة خبراء وتطبيق المنطق. كان أحد الإنجازات الرئيسية في هذه الفترة هو تطوير أنظمة الخبراء، التي احتفظت بمعرفة الخبراء وتمثيلها بلغة رمزية (هارمون وكينغ، 1985). ركزت أبحاث أنظمة الخبراء على أدوات اكتساب المعرفة للمساعدة في أتمتة عملية تصميم وتصحيح وصيانة القواعد التي حددها الخبراء. ومع ذلك، كان من الصعب الحصول على خبرة الخبراء في المجال، حيث كانت المعرفة الخبرائية تتغير باستمرار، بسبب التغيرات في اللوائح والمعايير البيئية. تدهورت الفترة ببطء إلى شتاء الذكاء الاصطناعي.
تعلم الآلة والتعلم العميق (التسعينات – 2020). غيرت التطبيقات الحديثة لتعلم الآلة والتعلم العميق ممارسة البحث والتطوير وبدأت في تسريع النمو عبر مجالات الأعمال والعلوم والهندسة المختلفة. زادت سرعة الابتكار في الذكاء الاصطناعي، مدفوعة بالبيانات الضخمة، وتوسع قوة الحوسبة، والخوارزميات والأساليب الجديدة. كان أحد الإنجازات الكبرى هو ImageNet في عام 2009 (دينغ وآخرون، 2009)، الذي احتوى على ملايين الصور المعلّمة في أكثر من 20,000 فئة وكان حاسمًا في إثبات شرعية استخدام نماذج ما قبل التدريب لتدريب الشبكات العصبية العميقة الكبيرة. منذ ذلك الحين، تم إحراز تقدم كبير في إنشاء الشبكات العصبية العميقة، مما ساهم في التقدم في تطبيقات مثل رؤية الكمبيوتر، ومعالجة اللغة الطبيعية (NLP) والروبوتات.
في عام 2017، طور باحثو جوجل بنية المحولات (فاسواني وآخرون، 2017)، التي أصبحت فيما بعد معتمدة على نطاق واسع في تطوير نماذج اللغة الكبيرة، نظرًا لقدرتها على معالجة اللغة الطبيعية. تستخدم البنية التعلم الذاتي، مع كمية كبيرة من نصوص الإنترنت المدخلة في النموذج لتدريبه. في قلب البنية هو بناء الانتباه الذاتي من الجمل (فاسواني وآخرون، 2017) الذي يسجل المعلومات المهمة ضمن سياق الجملة حول كيفية ارتباط الكلمات ببعضها البعض بطرق مختلفة. لقد تحسنت أداء نماذج ما قبل التدريب المعتمدة على المحولات بشكل كبير، مع تمكين نماذج اللغة الكبيرة بشكل مميز من الانتقال من الرمزية إلى الاتصال.
اكتسب الاتصال زخمًا في التسعينات، مما أدى تدريجيًا إلى استبدال الرمزية، التي كانت تعتمد على نموذج العقل البشري لتحقيق استدلال مشابه لذلك الذي يقوم به البشر (تشانغ وآخرون، 2023). يعتمد الاتصال على نمذجة الاتصال والأنظمة التي تحتوي على شبكات كبيرة من معالجات بسيطة للغاية، مترابطة بشكل كبير، وتعمل بالتوازي، كما هو موجود عادة في تقنيات التعلم العميق المختلفة (سمولينسكي، 1987). على الرغم من أن روزنبلات (1961) بنى بيرسيبترون في عام 1957 – النموذج الأولي لشبكة عصبية اصطناعية (ANN) – استنادًا إلى فكرة الاتصال، إلا أن الشبكات العصبية لم تحقق تطبيقات واسعة حتى التسعينات (تشانغ وآخرون، 2023).
تركز الإنجازات الحالية في الذكاء الاصطناعي على نماذج اللغة الكبيرة والخوارزميات المرتبطة بها. تحاكي نماذج اللغة الكبيرة التفكير البشري من خلال الاتصال، الذي يتجذر في تعلم الآلة عبر الشبكات العصبية، وموارد البيانات الضخمة المجمعة من مجموعة متنوعة من القنوات الرقمية. يتطلب تطوير تقنيات ذكاء اصطناعي جديدة تكون آمنة وموثوقة وقابلة للتوسع، نظرية ذكاء اصطناعي جديدة، قابلة للتفسير وقوية، مما يؤدي إلى تطوير جيل ثالث من الذكاء الاصطناعي من خلال دمج paradigms المختلفة للذكاء الاصطناعي (تشانغ وآخرون، 2023). يلخص الجدول 2 التقدم الكبير في الذكاء الاصطناعي الذي أدى إلى نماذج اللغة الكبيرة قبل عام 2020. توضح الجداول 1 و 2 المعالم المختارة في تطور تقنيات الذكاء الاصطناعي التي أدت إلى أحدث طفرة في الذكاء الاصطناعي التوليدي ونماذج اللغة الكبيرة. من المهم أن يكون باحثو نظم المعلومات على دراية بأن تقنية الذكاء الاصطناعي المبكرة (قبل 2017) كانت تركز بشكل أساسي على تطبيقات محددة مثل الألعاب والروبوتات. بينما تحولت تقنية الذكاء الاصطناعي الأكثر حداثة (منذ 2017) إلى معالجة اللغة الطبيعية والتطبيقات التوليدية. ومع ذلك، فقد تطورت المفاهيم والخوارزميات لتقنية الذكاء الاصطناعي بعد عام 2017 من الجهود الابتكارية التي سبقتها.

2.3 تطور الذكاء الاصطناعي وتأثيره على أبحاث نظم المعلومات

لقد أثر تطور هذه التقنيات في الذكاء الاصطناعي بشكل كبير على كل من الأنظمة الاجتماعية التقنية وأبحاث نظم المعلومات (IS). كانت النسخ السابقة من أدوات الذكاء الاصطناعي ضيقة ومحددة المهام، مما عزز بشكل أساسي قدرات الأنظمة لعمليات اتخاذ القرار المحددة. بالنسبة لباحثي نظم المعلومات، قدم هذا رؤى قيمة حول كيفية تكيف المنظمات مع التغيرات التكنولوجية واستغلال البيانات لفهم سياقات الأعمال المختلفة. ومع ذلك، كانت هذه الأنظمة السابقة محدودة في نطاقها وتعقيدها وغالبًا ما تفتقر إلى القدرة على التفاعل مع العمليات التجارية الأكثر تعقيدًا.
على النقيض من ذلك، وسعت ظهور GenAI تأثيرها عبر مجالات اجتماعية وتنظيمية متنوعة. GenAI أكثر قابلية للتكيف وواعية بالسياق، مما يمكّن الشركات من أتمتة المهام الأكثر تعقيدًا وتسهيل رؤى أعمق في عمليات اتخاذ القرار. يبرز هذا التحول الدور الحاسم لأبحاث نظم المعلومات.
الجدول 2 ملخص للتقدمات الحديثة في الذكاء الاصطناعي
السنة أداة الذكاء الاصطناعي أهمية الذكاء الاصطناعي المرجع
1989 خوارزمية Q-learning “التعلم من المكافآت المتأخرة” يحسن التعلم المعزز واتكينز (1989)
1993 حل “مهمة التعلم العميق جدًا” العالم حل المهمة بأكثر من 1,000 طبقة في الشبكة العصبية المتكررة (RNN) شمديهوبر (1993)
1995 نجاح SVM تم تطبيق آلات الدعم المتجهة على تصنيف النصوص، والتعرف على الأحرف المكتوبة بخط اليد، وتصنيف الصور كورتيس وفابنيك (1995)
1997 بنية LSTM حسنت بنية الذاكرة طويلة وقصيرة المدى (LSTM) RNN من خلال القضاء على مشكلة الاعتماد على المدى الطويل هوكرايتر وشمديهوبر (1997)
1998 تحسين التعلم القائم على التدرج يجمع بين خوارزمية التدرج العشوائي مع خوارزمية الانتشار العكسي ليكون وآخرون (1998)
2002 TD-Gammon تطابق أفضل لاعب يجمع بين الشبكات العصبية والتعلم المعزز (RL) مع طريقة اللعب الذاتي تيسارو (2002)؛ بلي وآخرون (2003)
2005 فاز الروبوت من جامعة ستانفورد بجائزة قاد 131 ميلاً بشكل مستقل على مسار صحراوي غير متكرر في تحدي داربا الكبير ثرون وآخرون (2006)
2011 فاز IBM Watson بجائزة Jeopardy Watson هو نظام أسئلة وأجوبة يجمع بين التعرف على الصوت، وتوليد الصوت، واسترجاع المعلومات، من بين أمور أخرى فيروتشي (2012)
2012 AlexNet فاز بمسابقة ImageNet، مما قد يشير إلى نقطة تحول في التعلم العميق كريزيفسكي وآخرون (2012)؛ كريزيفسكي وآخرون (2017)
2014 الشبكات التوليدية المتعارضة (GANs) هياكل الشبكات العصبية العميقة المكونة من شبكتين؛ تتعلم تقليد توزيع البيانات لتوليد محتوى مثل الصور والموسيقى والكلام، إلخ جودفيلو وآخرون (2014)؛ كينغما وويلينغ (2014)
2017 المحول هيكل DL يعتمد على آلية الانتباه الذاتي، مهم في نمذجة اللغة والترجمة الآلية والإجابة على الأسئلة فاسواني وآخرون (2017)
2018 OpenAI Five هزمت فريقًا بشريًا في Dota 2، لعبة فيديو معقدة؛ أكثر واقعية من الشطرنج أو Go باتشوك وآخرون (2018)
2019 GPT-2 نموذج لغة غير خاضع للإشراف على نطاق واسع؛ يمكنه توليد فقرات نصية متماسكة، وفهم القراءة، والترجمة الآلية، والإجابة على الأسئلة، والتلخيص رادفورد وآخرون (2019)
في اعتماد واستخدام GenAI. كما هو موضح في الشكل 1، يواجه الباحثون في نظم المعلومات الآن تحدي استكشاف كيف تعيد هذه الأنظمة المتقدمة تشكيل جميع مجالات الأنشطة البشرية، من عمليات العمل التنظيمية إلى التعاون بين الإنسان والذكاء الاصطناعي، مع تحديد النتائج غير المتوقعة التي قد تنشأ من هذا التكامل. تستكشف هذه الورقة هذه التطورات وتحدد أسئلة بحثية محتملة ضمن نظم المعلومات.

3 الإطار النظري لفهم الذكاء الاصطناعي التوليدي

في هذا القسم نستخلص إطارًا نظريًا لـ GenAI. الإطار مصمم للمساعدة في فهم GenAI ودعم البحث المقترح.

3.1 تصور الذكاء الاصطناعي التوليدي

لفهم الإمكانيات والتحديات، من المهم تقدير الآليات لمعالجة البيانات المتأصلة في GenAI. لكشف هذه الآليات، نفحص ChatGPT، وهو نوع رئيسي ومثال على GenAI. لاحظ أننا نستخدم ChatGPT كحالة من LLMs ولكن المناقشة ذات صلة بـ LLMs الأخرى. يقدم الشكل 2 تصورًا لـ ChatGPT، الذي سنفصل فيه لاحقًا.
كما هو موضح في الشكل 2، GenAI، مثل ChatGPT هو نظام متقدم للإجابة على الأسئلة يعتمد على نماذج حسابية متعددة وبيانات ضخمة. تتبنى ChatGPT و LLMs الأخرى نهجًا مركزيًا للبيانات من خلال تدريب نماذجها لتوليد استجابات شبيهة بالبشر بناءً على مطالبات المستخدم بلغة بشرية. تشمل النماذج الأساسية المحول (فاسواني وآخرون، 2017)، الانتباه (لونغ وآخرون، 2015)، GAN
الشكل 2 الذكاء الاصطناعي التوليدي كنظام متقدم للإجابة على الأسئلة
(جودفيلو وآخرون، 2014، 2020)، نماذج المكافآت (يو وآخرون، 2020)، و RLHF (لي وآخرون، 2023)، والتي تُستخدم في توليد الاستجابات من النصوص في مستودعات البيانات الضخمة (جودفيلو وآخرون، 2014، 2020؛ لي وآخرون، 2023؛ لونغ وآخرون، 2015؛ فاسواني وآخرون، 2017؛ يو وآخرون، 2020). في هذا المعنى لتحويل النص، يُشار إلى LLMs على أنها GenAI. بالمقابل، في أنظمة الأسئلة والأجوبة التقليدية، قد تتطابق الإجابات المخزنة مسبقًا مع الكلمات الرئيسية في أسئلة المستخدم باستخدام بعض آليات البحث، وبالتالي تكون غير توليدية. تلخص الجدول 3 المفاهيم الرئيسية المتعلقة بنماذج اللغة الكبيرة وآلياتها.
نناقش الآن الجوانب الرئيسية لتقدمات GenAI كنوع فريد من تكنولوجيا المعلومات. لا يُقصد من هذا الاستكشاف أن يكون حاسمًا بل لتحفيز مزيد من المناقشة والبحث من منظور تطبيقي، بدلاً من المنظور الفني البحت.
النهج التنبؤي – ChatGPT هو نهج تنبؤي يعتمد على الشبكات العصبية العميقة. ينجح ChatGPT في مهام معالجة اللغة الطبيعية مثل الدردشة والترجمة وصنع القصص والترميز وغيرها من المهام المماثلة بعدة لغات. وذلك بسبب تمثيله
للغة الطبيعية البشرية في الشبكات العصبية العميقة مثل المحولات. كما هو الحال في العديد من تقنيات التعلم الآلي، يتم تدريب ChatGPT مسبقًا بناءً على النصوص الموجودة ثم يُستخدم للتنبؤ بالنتائج بناءً على نص المطالبة من المستخدم. يعني هذا النهج التنبؤي أن الاستجابات من ChatGPT احتمالية وبالتالي عرضة للخطأ بالضرورة. هذه الخاصية التنبؤية سهلة الفهم من قبل المحترفين الفنيين، ولكنها تتطلب شرحًا للمحترفين غير الفنيين.
النماذج الاتصالانية – ChatGPT توليدي لأنه ينشئ اتصالات جديدة بين الكلمات عبر نماذج اللغة الكبيرة. كما هو ملاحظ، تطور الذكاء الاصطناعي إلى نماذج اتصالانية بعد استكشاف النماذج الرمزية ثم حقق نجاحًا كبيرًا في معالجة اللغة الطبيعية عبر نماذج اللغة الكبيرة، مما أدى إلى ChatGPT وأدوات مماثلة (تشانغ وآخرون، 2023). تقوم الشبكات العصبية أساسًا بإنشاء اتصالات بسيطة وموزونة بين الجمل والكلمات بعد عمليات الترميز، مما يسمح بحسابات التعلم الآلي على مئات المليارات من المدخلات (مثل رموز النص). لم يكن من الممكن تحقيق نجاح ChatGPT لأن النماذج الرمزية البحتة يصعب توسيعها إلى مستويات البيانات الضخمة. هذا الاتصالانية
الجدول 3 المفاهيم الأساسية لنماذج اللغة الكبيرة
المفهوم التعريف السنة
GAN تقوم الشبكات التنافسية التوليدية (GAN) بتدريب نموذجين في وقت واحد: نموذج توليدي يلتقط توزيع البيانات ونموذج تمييزي يقدر احتمال أن عينة جاءت من بيانات التدريب 2014
الانتباه تُستخدم آلية الانتباه في التعلم الآلي ومعالجة اللغة الطبيعية لزيادة دقة النموذج من خلال التركيز على البيانات ذات الصلة لتمكين النموذج من التركيز على مجالات معينة من بيانات الإدخال، مما يمنح وزنًا أكبر للميزات الحاسمة وتجاهل غير المهمة 2015
المحول المحولات هي نوع من الشبكات العصبية مصممة للتعامل مع الاعتماديات بعيدة المدى في النص، والتقاط العلاقات بين الكلمات والسماح للنموذج بفهم السياق والمعنى عبر تسلسلات كبيرة من المدخلات. على سبيل المثال، العلاقة والملاءمة بين اللون والسماء والزرقاء في السؤال: “ما هو لون السماء؟” تؤدي إلى الناتج: “السماء زرقاء.” 2017
المكافأة نموذج المكافأة هو نهج في الذكاء الاصطناعي حيث يتلقى النموذج مكافأة أو درجة لاستجاباته للمطالبات المعطاة. تعمل هذه الإشارة المكافئة كتعزيز، موجهة نموذج الذكاء الاصطناعي لإنتاج النتائج المرغوبة 2020
RLHF في التعلم الآلي، التعلم المعزز من ردود الفعل البشرية (RLHF) هو تقنية لمواءمة وكيل ذكي مع تفضيلات البشر 2023
يمكن أن تربط ChatGPT بشكل طبيعي مع التطبيقات التجارية عبر روابط إضافية بين الإنسان والآلة لتحقيق الكفاءة والدقة. ومع ذلك، لا تتطلب الاتصالانية في ChatGPT معرفة عميقة بموضوع ما أو الحدس أو التفكير السليم للبشر. وهذا يترك مجالًا للتحسينات في التفكير والإبداع. وبالتالي، على الرغم من أن LLMs أظهرت أنها تنتج نصوصًا متماسكة وشبيهة بالبشر، إلا أنها لا ‘تفهم’ اللغة بالمعنى البشري ولا يمكنها التمييز بين المعلومات الواقعية وغير الواقعية (سابا، 2023).
القدرات العامة – ChatGPT عامة لأن النماذج الاتصالانية لا تميز بين القطاعات التجارية أو مجالات المعرفة. يمكن أن يخدم ChatGPT واحد مليارات المستخدمين للأسئلة والأجوبة، دون أن يكون مقيدًا بسياقات محددة. هذه الميزة هي نتيجة لنماذج اللغة الكبيرة والشبكات العصبية العملاقة من الاتصالات. ومع ذلك، فإن هذه القدرة، التي تعتمد على مجموعة بيانات كبيرة، تعاني أيضًا من انخفاض الدقة وإمكانية الهلوسة (أي، اختراع مخرجات غير منطقية).
التكامل مع محركات البحث – تكمل نتائج ChatGPT تلك من محركات البحث لأن الأولى تنبؤية؛ والأخيرة هي ببساطة مجموعة من الوثائق الموجودة ذات الصلة باستعلام المستخدم. على هذا النحو، يمكن اعتبار ChatGPT ملخصًا للوثائق الموجودة التي ترتبط بشكل أكثر دقة باستعلام المستخدم، بينما توفر محركات البحث ببساطة قائمة بنتائج البحث للمستخدم. بطريقة ما، تقدم ChatGPT نتيجة أولية للمستخدم، الذي قد يعالج الإجابة من ChatGPT بشكل أكبر. هناك علاقة مثلثية بين المستخدم وChatGPT ومحرك البحث حيث قد يحدد المستخدم كيفية استخدام ومعالجة النتائج من ChatGPT ومحرك البحث. يجب أن يكون هذا موضوع بحث مثمر للغاية حول كيفية إدارة نظام معالجة المعرفة هذا. لن تحل العلاقة المثلثية محل محركات البحث لأن نتائج ChatGPT قد لا تلبي جميع احتياجات المستخدمين، على الرغم من أن بعض المستخدمين قد يقبلون النتائج، دون استخدام محرك البحث مرة أخرى (دوبين وآخرون، 2023).
قدرة الفهم لدى ChatGPT – كيفية فهم ChatGPT للنصوص مقارنة بالبشر هو موضوع بحث مهم. يتعلم البشر طوال حياتهم، بدءًا من الطفولة، ويغنون عقولهم بمعرفة جديدة دون توقف طوال حياتهم. يعبر الناس عن الأفكار، ويتبادلون الآراء، ويخلقون مفاهيم وكلمات ونظريات جديدة. يمكن لـ ChatGPT معالجة كميات هائلة من النصوص في الوثائق الموجودة وتوليد نصوص جديدة بناءً على طلبات المستخدمين. ومع ذلك، لوحظ أن الذكاء الاصطناعي لا يمكنه فهم معنى السياق بالمعنى البشري لأنه لا يتعامل مع المعنى الدلالي العميق للكلمات التي يعالجها، مما يوضح أن هناك فجوة كبيرة جدًا لا تزال قائمة بين الذكاء الاصطناعي-
ليس من المعقول ببساطة أن نؤكد أن ChatGPT يمكنه فهم أو استيعاب الوثائق بالمعنى البشري. هناك فرق كبير بين إنتاج نص جديد وابتكار أفكار جديدة.

3.2 نظرية الأنظمة: شرح الذكاء الاصطناعي التقليدي مقابل الذكاء الاصطناعي التوليدي

لفهم ما يجعل الذكاء الاصطناعي التوليدي محوريًا، نقوم بتطوير إطار نظري يأخذ في الاعتبار بنية الذكاء الاصطناعي التوليدي ويستند إلى نظرية الأنظمة وغيرها من الحجج النظرية (مثل النظرية اللغوية). تساعد هذه الأسس في تحديد الخصائص الفريدة للذكاء الاصطناعي التوليدي وسلوكه.
لقد استندت العديد من الجهود لفهم وتصميم تكنولوجيا المعلومات إلى نظرية الأنظمة (Chatterjee et al., 2021). تشمل الأمثلة التحقيقات في الذكاء الاصطناعي بشكل عام (Skyttner, 2001)، والثقة في الذكاء الاصطناعي (Lukyanenko et al., 2022a)، وتفاعلات الإنسان مع الروبوتات (Lima & Custodio, 2004). كما تم اقتراح نظرية الأنظمة سابقًا لفهم ChatGPT (Dwivedi et al., 2023). لذلك، نعتمد على نظرية الأنظمة لفهم GenAI ونؤكد أن أنواعًا مختلفة من GenAI (مثل تلك المستندة إلى نماذج اللغة الكبيرة) هي أنواع من الأنظمة التوليدية.
نظرية الأنظمة تغطي النظريات ذات الصلة والمتداخلة التي تتعامل مع طبيعة الأنظمة وتفاعلاتها واستخداماتها. من أجل الأسس العامة، نبني على أعمال أكوف (1971) وآخرين، ونكملها بأبحاث السوسيولوجيا التقنية التي تدرس الأنظمة ضمن السياقات الاجتماعية (تشاتيرجي وآخرون، 2021).
النظام هو مفهوم علمي واجتماعي أساسي: “كيان يتكون من عنصرين على الأقل … كل عنصر من عناصر النظام مرتبط بكل عنصر آخر، بشكل مباشر أو غير مباشر. لا يوجد مجموعة فرعية من العناصر غير مرتبطة بأي مجموعة فرعية أخرى” (أكاوف، 1971، ص. 662). يحتوي GenAI على مكونات مترابطة عديدة. عند النظر في LLM المعتمد على المحولات، هناك LLMs للترميز والترميز العكسي لعملية منطق أنظمة الإدخال والإخراج. يتطلب تصور GenAI كنظم: 1) تحليل الخصائص الرئيسية للنظم، مثل الظهور؛ و 2) التحقيق في كيفية أن تصبح هذه التكنولوجيا مكونًا من نظم أوسع، خاصة النظم الاجتماعية التقنية.
تمتلك الأنظمة نوعين من الخصائص (بونج، 1979، 2018): خصائص الأجزاء (المعروفة بالوراثية) وخصائص الأنظمة نفسها (المعروفة بالناشئة). على سبيل المثال، كتلة المركبة هي خاصية وراثية، وهي مجموع كتل مكوناتها. في سياق الذكاء الاصطناعي، مكونات الشبكة العصبية هي العقد المتصلة. الخاصية الوراثية هي معامل الاتصال بين العقد.
بالإضافة إلى ذلك، تمتلك الأنظمة خصائص ناشئة، وهي خصائص لنظام كامل، بدلاً من أن تكون لأي من مكوناته.
تظهر هذه الخصائص عندما تصبح المكونات جزءًا من الكل وتبدأ في التفاعل مع بعضها البعض بطريقة معينة. نظرًا لأن أي مكون لا يمتلك خاصية ناشئة، فإن الخصائص الناشئة غالبًا ما لا يمكن اشتقاقها من معرفة خصائص المكونات. على سبيل المثال، التضامن هو خاصية ناشئة لحزب سياسي؛ لا يمتلك أي عضو في الحزب هذه الخاصية. تعتمد هذه الخاصية، ليس فقط على معتقدات وسلوكيات الأعضاء الأفراد (بالإضافة إلى عوامل خارجية)، ولكن أيضًا على التاريخ والديناميات والتفاعلات بين أعضائه. وبالمثل، فإن الشفافية هي الخاصية الناشئة للشبكة العصبية بأكملها. بينما قد تكون مكونات الشبكة العصبية الفردية مفهومة للإنسان، عندما يتم تجميع هذه المكونات معًا، قد تفتقر إلى الشفافية. وبالمثل، فإن المخرجات التي تبدو بشرية من ChatGPT هي ناشئة من البنية المعمارية الخاصة بالاتصال والشبكات اللغوية الطبيعية لـ GenAI، والتي، إذا تم فصلها واستخدامها بشكل منفصل، ستفشل في تقديم أنواع مماثلة من المخرجات.
تشكل الخصائص الناشئة أيضًا السلوك الناشئ للأنظمة، على غرار سلوك تجمع الأسماك. حيث يتم تشكيل السلوك الناشئ بواسطة الخصائص الناشئة الغامضة، تُشكل الأنظمة الناشئة بواسطة خصائصها الناشئة، ومن ثم سلوكها. تؤدي الناشئة إلى تجديد نوعي، وربما وجودي (بدو & همفريز، 2008). تخلق الناشئة في الأنظمة واقعًا جديدًا، لا يوجد على مستوى مكوناتها. غالبًا ما يمكن للأنظمة المعتمدة على الذكاء الاصطناعي تقديم هذا النوع من الخصائص. على سبيل المثال، يمكن لـ GAN داخل GenAI، استنادًا إلى مفهوم التقليد الطبيعي، توليد صور تُعتبر جديدة.
بشكل عام، كلما كان النظام أكثر تعقيدًا، كان من الصعب فهمه والتنبؤ بسلوكه. تفهم نظرية الأنظمة التعقيد على أنه مُدرك وفعلي. التعقيد المدرك هو تفسير الإنسان وتصوره لنظام ما على أنه معقد (Li & Wieringa، 2000؛ Schlindwein & Ison، 2004). تتأثر إدراك التعقيد جزئيًا وبشكل إيجابي بالتعقيد الفعلي. يمكن فهم التعقيد الفعلي على أنه عدد الأجزاء المكونة، إلى جانب الطريقة التي يتم بها هيكلة هذه الأجزاء وتفاعلها مع بعضها البعض (Lukyanenko et al.، 2022b). بالإضافة إلى المفاهيم النظامية العامة، تنطبق اعتبارات محددة على أنواع معينة من الأنظمة (مثل: هادفة، تكيفية، عضوية، صناعية، ذاتية التنظيم، ذاتية الانعكاس، ملموسة، مفاهيمية) (Ackoff & Emery، 2005). يتيح تسليط الضوء على الفروق فهمًا أعمق لـ GenAI.
علاوة على ذلك، نلاحظ التمييز بين الأنظمة الملموسة والمفاهيمية. الأنظمة الملموسة هي أنظمة تتكون من مكونات مادية (مثل: الفيزيائية) (بونج، 1996). يمكن أن تتفاعل الأنظمة الملموسة مباشرة مع أنظمة أخرى وتتغير مع احتواء مكوناتها المادية على طاقة، يمكن أن تستجيب وتثير التغيير. الحواسيب والبشر أنظمة ملموسة. المنظمات، التي هي اجتماعية
الأنظمة أيضًا ملموسة، لأنها مصنوعة من مكونات ملموسة – البشر وقطعهم الفنية (لوهمان، 1995).
الأنظمة المفاهيمية هي أفكار مجردة مرتبطة معًا في عقل نظام ملموس (مثل: الكائن البشري) من خلال قواعد عقلية، مثل المنطق (بونج، 1996). المعادلات، النظريات، التسلسلات الهرمية، الأطر، قواعد اللغة، السرد النصي المبني منطقيًا (مثل: المقالات، الفقرات) هي أنظمة مفاهيمية. على عكس الأنظمة الملموسة، لا تحتوي الأنظمة المفاهيمية على طاقة وتُنتج بواسطة الأنظمة الملموسة، التي تستهلك الطاقة لإنشاء وتخزين وتعديل والتواصل مع هذه الأنظمة.
نوع آخر مهم من الأنظمة هو النظام التكيفي. على عكس الأنظمة الثابتة أو الصارمة، فإن الأنظمة التكيفية قادرة على الاستجابة للتغيرات البيئية من خلال إعادة تكوين حالاتهم الداخلية. مجموعة فرعية من هذه الأنظمة هي الأنظمة التكيفية المعقدة، والتي تُعرف بأنها “أنظمة تتكون من وكلاء يتفاعلون ويتم وصفهم من حيث القواعد. يتكيف الوكلاء من خلال تغيير قواعدهم مع تراكم الخبرة” (هولاند، 1992، ص. 10). تشمل الأنظمة التكيفية المعقدة الكائنات الطبيعية والأنظمة الاصطناعية، مثل الآلات المعقدة، بما في ذلك تلك المعتمدة على الذكاء الاصطناعي. تعتمد هذه الأنظمة عادةً على حلقات التغذية الراجعة، حيث تصبح مخرجات النظام مدخلاته، وبالتالي يمكن أن تعدل أو تضخم سلوك النظام. نتيجة مهمة للتكيف هي زيادة الصعوبة في توقع وتنبؤ سلوك هذه الأنظمة (هولاند، 1992).
الهدف من عملية GenAI هو إنتاج مخرجات جديدة ومعقدة ومكتفية ذاتيًا، على عكس التعلم الآلي التقليدي، الذي يركز بشكل أساسي على تعلم حدود القرار بناءً على الأنماط المستخرجة من البيانات (والترز وموركو، 2020). هناك فرق نوعي عميق بين GenAI والأنظمة التقليدية التمييزية. تقنيات الذكاء الاصطناعي التقليدية هي نماذج قرار في الأساس، هياكل رياضية تسعى لربط المدخلات بالمخرجات بطريقة مباشرة نسبيًا. على الرغم من كونها غامضة ومعقدة، فإن مبادئ تشغيلها تشبه تلك الخاصة بشريط قياس أو آلة حاسبة. غالبًا ما تكون مساحة الحل أو المخرجات محدودة، مما يجعلها تركز على مهام محددة، وبالتالي تعتبر ذكاءً ضيقًا.
على النقيض من ذلك، فإن GenAI، الذي يعتمد أيضًا على مليارات التكرارات من التدريب، ويعبر عنه كهياكل رياضية معقدة، يحدد المخرجات بطريقة أقل مباشرة بكثير. بدلاً من القواعد التي تربط المدخلات بالمخرجات، فإن GenAI هو مجموعة من المعلمات التي توجه تطوير مخرجات مكتفية ذاتيًا بناءً على طلب المدخلات. تشبيه مفيد هو أنظمة اللغة الطبيعية.
وجهة نظر سائدة في علم اللغة هي أن اللغات الطبيعية موجهة بواسطة القواعد النحوية العالمية (UG) – المبادئ الغريزية تتجسد عندما يتعلم المتحدث لغة معينة ويستخدمها (تشومسكي، 1986). تحدد UG المبادئ العامة والمعلمات (على سبيل المثال، يمكن تحديد كل كلمة بفئة لغوية). ومن ثم، فإن UG هو لغة ميتا. نتيجة مبادئ UG والمعلمات هي
إنتاجية اللغات البشرية. من أجل هدف، يمكن إنتاج تعبيرات مختلفة بلا حدود. أيضًا، وفقًا لـ UG، فإن مخرجات اللغة الطبيعية متماسكة ومكتفية ذاتيًا. على سبيل المثال، الجملة، الفقرة، أو المقالة جميعها لها هيكل داخلي متسق ومتناسق، موجه بمبادئ ومعلمات UG.
بنفس الطريقة التي تعمل بها اللغة الطبيعية وأنظمة الإنتاج الأخرى (مثل نظام الأعداد)، يمكن لـ GenAI إنتاج مخرجات غير محدودة محتملة، بناءً على مدخلات مشابهة (تحت الافتراض الشائع بأن معلمة درجة الحرارة فوق الصفر). بينما ترتبط هذه المخرجات في النهاية بالبيانات، فإن الجدة تكمن في الروابط الجديدة وتحويل هذه البيانات. نظرًا لمجموعة المعلمات ومصادر البيانات الكبيرة جدًا، فإنه من المستحيل تقريبًا التنبؤ بكل طريقة ممكنة قد يربط بها GenAI مدخلًا بمخرج. على الرغم من أن مقارنة الأدوات، مثل ChatGPT، بالإبداع البشري لا تزال مثيرة للجدل، فإن GenAI يعد بزيادة كبيرة في مستوى تجديد المخرجات مقارنة بمعظم أشكال الذكاء الاصطناعي الأخرى.
بينما قد يوحي هذا بأنه من المستحيل السيطرة تمامًا على GenAI، فإن نظرية الأنظمة ونظرية اللغة تقدم رؤى حول كيفية التنبؤ وإدارة هذه الأنواع من التقنيات. على وجه التحديد، فإن المبادئ والمعلمات الخاصة بـ GenAI مهمة. يتم تعلم الهياكل اللغوية للغة بواسطة النظام، ويتم رسمها إلى لغة أخرى أو صور، وتطبق في النهاية لمهام مختلفة؛ على سبيل المثال، تعليمات نصية لتوليد نص برمجي بلغة بايثون.
نظرًا لقدرتها الإنتاجية، فإن أنظمة الذكاء الاصطناعي، مثل ChatGPT، تظهر تعقيدًا أكبر بكثير من أنظمة التعلم الآلي التقليدية (بندر وآخرون، 2021). هذه الأنظمة قادرة أيضًا على ربط ودمج أنواع البيانات التي تتجاوز النص (مثل الصورة والكلام)، ونتيجة لذلك، من المتوقع أن يكون لديها عدد أكبر بكثير من الخصائص الناشئة، مما يؤدي إلى مجموعة متنوعة من السلوكيات الناشئة. وفقًا لنظرية الأنظمة، فإن قدرة أنظمة الذكاء الاصطناعي على التنبؤ بالسلوك (علاقة المدخلات بالمخرجات) في الأنظمة الإنتاجية، تقل بشكل كبير مقارنة بأنظمة التعلم الآلي التقليدية.
تسمح نظرية الأنظمة برؤى ملحوظة أخرى حول الأنظمة الإنتاجية. ليس فقط أن GenAI هو نظام بحد ذاته، ولكن مدخلاته ومخرجاته غالبًا ما تكون أنظمة (ليس فقط مكونات النظام). على سبيل المثال، أنظمة تحويل النص إلى صورة، مثل DALL-E، تجمع بين LLMs ونماذج الانتشار. الأول يبني فهم الصور المرتبطة بالنصوص المدخلة؛ والأخير يتبع نموذجًا احتماليًا لتوليد صور مشابهة للنصوص المدخلة. تم تطبيق النموذجين سابقًا بشكل مستقل في أنظمة تركز على المهام المعنية.
كانت الأجيال السابقة من الذكاء الاصطناعي تعمل فقط بمكونات النظام كمدخلاتها ومخرجاتها. كان نموذج التعلم الآلي النموذجي يقبل مدخلات بتنسيق محدد مسبقًا (يتم التحكم فيه خارجيًا من خلال واجهة المستخدم)، وينتج مخرجات، والتي، بدقة، كانت
غير قابلة للاستخدام بمفردها، ولكنها تتطلب تفسيرًا ودمجًا في أنظمة أكبر (مفاهيمية أو ملموسة). على سبيل المثال، يمكن أن ينتج نموذج التعلم الآلي مخاطر ائتمانية لعميل، بناءً على متجه ميزات محدد مسبقًا يتوافق مع معلمات عميل معين. ومع ذلك، يجب تفسير النتائج مثل , ويجب دمجها في نظام مفاهيمي أكبر.
يمكن للذكاء الاصطناعي الإنتاجي، مع ذلك، قبول وتوليد أنظمة مكونة، فضلاً عن أنظمة بحد ذاتها. في الواقع، عندما يكتب ChatGPT مقالًا، أو ينشئ نصًا برمجيًا بلغة بايثون، أو يجيب على سؤال؛ عندما يرسم DALL-E أو MidJourney صورة؛ أو عندما ينشئ MusicLM موسيقى، تكون النتيجة نظامًا مفاهيميًا مستقلًا. لم يعد هذا النظام بحاجة إلى أن يتم تضمينه في نظام آخر ليكون قابلاً للاستخدام. يمكن استخدامه مباشرة، على سبيل المثال، للاستماع إلى الموسيقى، أو الاستمتاع بصورة، أو تقديم مقال كواجب دراسي (وهو، بالطبع، يطرح أسئلة أخلاقية وتربوية أساسية).
أخيرًا، فإن القدرة على إنتاج مجموعة واسعة من المخرجات، سواء كانت أنظمة مكونة أو أنظمة بحد ذاتها، تسمح بتنوع أكبر في استخدام GenAI داخل المنظمات. مثل الذكاء الاصطناعي السابق، يمكن تضمين GenAI في تقنيات أخرى من خلال استغلال مخرجات نظام مكوناته من أجل اشتقاق نظام أوسع متماسك. على سبيل المثال، يمكن لنظام دعم القرار الموجود داخل منظمة ما استشارة GenAI لبعض المهام وتقديم مخرج لمستخدم يعتمد جزئيًا على ردود من تقنية GenAI. بالإضافة إلى ذلك، يمكن دمج GenAI بشكل أكثر مباشرة داخل الهياكل التنظيمية، لأنه ينتج أنظمة مفاهيمية مستقلة. على سبيل المثال، يمكن لموظفي المنظمة استخدام هذه التقنية لإعداد تقارير، كتابة مقالات، أو إجراء مراجعة لمجال ما. نظرًا لهذه القدرات المتنوعة، فإن سؤالًا مهمًا هو كيفية الاستفادة بشكل مناسب من هذه المرونة النظامية لـ GenAI.
استنادًا إلى نظرية الأنظمة وعلم اللغة الموضحة أعلاه، نقترح أن GenAI لديه ثلاث خصائص إنتاجية تميزه عن تقنيات التعلم الآلي والذكاء الاصطناعي في الماضي.
  • الظهور القوي. بينما قد تمتلك أنظمة الذكاء الاصطناعي السابقة خصائص ناشئة، فإن GenAI لديه قدرة ملحوظة على التصرف بطريقة لا يمكن اشتقاقها مباشرة من خصائص مكوناته وتكون بعيدة جدًا عنها. وذلك لأن نتائج GenAI هي نتيجة تحولات من المطالبات بالاقتران مع المعرفة المعقدة لنظام GenAI. يمكن أن يكون للظهور القوي عواقب إيجابية (مثل القدرة على إنتاج محتوى إبداعي) وسلبية (مثل صعوبة السيطرة على GenAI وضمان عدم إيذائه أو الإضرار بالناس).
  • الجدة الإنتاجية. متجذرة في الظهور القوي هي الجدة الإنتاجية. لدى الذكاء الاصطناعي الإنتاجي القدرة على إنتاج مخرجات متوقعة وغير متوقعة، بناءً على مدخل معين.
    المخرجات هي منتجات لمليارات من المعلمات المضبوطة من خلال مليارات من التكرارات على بيانات تدريب كبيرة. بينما ترتبط هذه المخرجات في النهاية ببيانات التدريب، تكمن الجدة في طرق جديدة لتحويل البيانات وتحديد الأنماط غير المرئية.
  • المدخلات والمخرجات النظامية. تمتلك الذكاء الاصطناعي التوليدي القدرة على قبول وإنتاج مخرجات متماسكة ومكتفية ذاتيًا (مثل الردود المكتفية ذاتيًا، المقالات، الصور، الرسوم المتحركة، الموسيقى). بشكل فعال، ينتج الذكاء الاصطناعي التوليدي أنظمة مفاهيمية مكتفية ذاتيًا (على عكس مقتطفات من قواعد القرار)، جنبًا إلى جنب مع مكونات النظام. بمعنى آخر، يتم رسم الأنظمة كمخرجات من مدخلات النظام بناءً على خوارزميات التحويل المتطورة.
تؤدي هذه الخصائص الثلاث إلى فرص جديدة للتفاعل بين الإنسان والكمبيوتر. تعيد آثارها التنظيمية تشكيل الإمكانيات لأبحاث نظم المعلومات بطرق غير مسبوقة في تاريخ نظم المعلومات.

3.3 المنظور الاجتماعي التقني

تركز الأبحاث في نظم المعلومات بشكل كبير على القضايا الاجتماعية التقنية حيث تعتبر المكونات التقنية في سياق أنظمة اجتماعية أوسع (Sarker et al., 2019). نظرًا لأن المنظور الاجتماعي التقني ينشأ أساسًا من نظرية الأنظمة، استخدمنا لفهم طبيعة الذكاء الاصطناعي التوليدي، من المعقول أن نعتبر الذكاء الاصطناعي التوليدي من المنظور الاجتماعي التقني.
تقدم الشكل 3 إطارًا للذكاء الاصطناعي التوليدي كنظام اجتماعي تقني. نظرًا لأنه يعتمد على منظور الأنظمة ويتبع المنظور الاجتماعي التقني، نقترح أن أبحاث نظم المعلومات يجب أن تأخذ في الاعتبار الخصائص التقنية
لتكنولوجيا الذكاء الاصطناعي التوليدي في سياق أنظمة أخرى، ومنظمات، وأفراد، وعمليات. على وجه التحديد، الخصائص الأساسية المميزة للذكاء الاصطناعي التوليدي من الظهور القوي، والجدة، والمدخلات والمخرجات النظامية، تنتج تحديات وفرص جديدة للتحقيق في تصميم واستخدام وتأثير هذه التكنولوجيا في السياقات التنظيمية والاجتماعية. يسمح هذا التحقيق لمستخدمي نظم المعلومات بتبني واستغلال المعرفة الواسعة الموجودة عن المنظمات والأنظمة ودمج هذه المعرفة مع الأسس الجديدة للذكاء الاصطناعي التوليدي. على سبيل المثال، مع الأبحاث الواسعة في نظم المعلومات حول التجارة الإلكترونية والتخصيص، فإن دمج الذكاء الاصطناعي التوليدي في منصات التجارة الإلكترونية (مثل محرك بحث Bing) يخلق فرصًا للتحقيق في: دور الذكاء الاصطناعي التوليدي في اعتماد المستخدم؛ الثقة تجاه تقنيات التجارة الإلكترونية؛ والقدرة على تسهيل التجارة وتبادل المعلومات. بسبب الخصائص الثلاثة الجديدة جوهريًا للذكاء الاصطناعي التوليدي (انظر الشكل 3)، يمكن ربط هذه التكنولوجيا بطرق جديدة مع الأنظمة الأوسع. يمكن أن يؤدي ذلك إلى فرصة بحث لدراسة تقاطع الجوانب التقنية للذكاء الاصطناعي التوليدي والطرق التي تنجح أو تفشل بها هذه القدرات في دعم وتعزيز الأنظمة التي تصبح جزءًا منها، أو تتفاعل معها.
يجب أن يتصرف النظام الاجتماعي التقني القائم على الذكاء الاصطناعي التوليدي بشكل متسق مع القيم الإنسانية (محاذاة الذكاء الاصطناعي) (Hagendorff & Fabi, 2022). تتضمن المحاذاة تحديد أولويات القيم الإنسانية الأساسية بدقة داخل نموذج اللغة الكبير. لتمثيل هذه القيم في نظام ذكاء اصطناعي، غالبًا ما يشارك البشر في ضبط نموذج اللغة الكبير باستخدام التعلم المعزز من ردود الفعل البشرية (RLHF). يجب على الباحثين في نظم المعلومات أن يأخذوا في الاعتبار الوجود الأوسع للنظام، بما في ذلك شروط الحدود التي تحدد وجوده وتأثيره في سياق فرد أو منظمة أو مجتمع.
الشكل 3 إطار للذكاء الاصطناعي التوليدي كنظام اجتماعي تقني
من خلال اعتماد المنظور الاجتماعي التقني، نقترح أن هناك آثار واسعة حول كيفية أن الذكاء الاصطناعي التوليدي لديه القدرة على تحويل المنظمات والمجتمع. الأول هو الاضطراب والتأثير على التطبيقات التي ستعمل على تحسين إنتاجية الإنسان والمنظمات بشكل كبير. مع تقدم تكنولوجيا الذكاء الاصطناعي التوليدي، ستساعد في تطوير حلول إبداعية ومبتكرة. ثانيًا، سيعزز نظام الذكاء الاصطناعي التوليدي العلاقات بين البشر والأنظمة التي تمتد عبر المساحات الشخصية والتجارية. يؤدي ذلك إلى قضايا بحث تركز على التعاون بين الإنسان والذكاء الاصطناعي. ثالثًا، الطبيعة المتنوعة للذكاء الاصطناعي التوليدي تقدم العديد من عدم اليقين، بعضها قد يوفر فرصًا جديدة ولكن أيضًا يشكل تهديدات للأفراد أو المنظمات أو المجتمع، والتي نعتبرها الجانب المظلم للذكاء الاصطناعي التوليدي. وبالتالي، تتطلب أنظمة الذكاء الاصطناعي التوليدي نموذج تصميم جديد. تشمل الموضوعات المحتملة لتصميم الأنظمة باستخدام الذكاء الاصطناعي التوليدي القضايا الاجتماعية المتعلقة بأتمتة الوظائف وإزاحة العمال البشر، مما قد يؤدي إلى مطالبات بمهارات جديدة وإعادة تدريب.

4 فرص البحث لأبحاث نظم المعلومات

تقترح هذه القسم مواضيع وأفكار بحثية للبحث المستقبلي حول الذكاء الاصطناعي التوليدي لنظم المعلومات والتخصصات ذات الصلة. تبني على الأسس النظرية الاجتماعية التقنية التي تعطي الأولوية لتقاطع الخصائص التقنية للذكاء الاصطناعي التوليدي وكياناته الاجتماعية (الأفراد، المنظمات، المجتمع).

4.1 الاضطراب والتأثير

من المحتمل أن يعزز الذكاء الاصطناعي التوليدي بشكل كبير إنتاجية الأفراد والمنظمات، مما يتسبب في اضطرابات في العديد من الصناعات. لدى باحثي نظم المعلومات سجل طويل من دراسة مثل هذه التأثيرات للتقنيات الجديدة.
تعزيز الإنتاجية وأتمتة المهام. تقوم خوارزميات التعلم الآلي بأتمتة مهام تصنيف البيانات والتنبؤ، مما يسمح بمعالجة وتحليل البيانات بشكل أكثر كفاءة. يمكن لأتمتة العمليات الروبوتية أتمتة العمليات التجارية الروتينية، مثل استخراج البيانات ودمج البيانات. نظرًا لأن الذكاء الاصطناعي التوليدي يمكنه معالجة اللغة، فإنه يمكن أن يقوم بأتمتة مجموعة أوسع من الوظائف، مما يؤدي إلى حالات استخدام جديدة. على سبيل المثال، يؤثر ChatGPT على الإعلانات بسبب مدى قدرته على إعداد رسائل التسويق والإعلانات بشكل جيد. يمكن أن تحرر أنظمة الذكاء الاصطناعي الناس لأداء أعمال أكثر إرضاءً أو
إبداعية. يمكن أن يؤدي ذلك إلى تحسين الكفاءة، وتسريع اتخاذ القرار، وزيادة الإنتاجية الإجمالية. لقد سعى باحثو نظم المعلومات تقليديًا إلى قياس قيمة التكنولوجيا في ظل ظروف مختلفة (Chang & Gurbaxani, 2012). وبالمثل، هناك حاجة لتقييم القيمة التي يقدمها الذكاء الاصطناعي التوليدي ضمن السياق التجاري الأكبر.
تحويل القوة العاملة والمنظمة. قد يتطلب اعتماد الذكاء الاصطناعي التوليدي تحويل القوة العاملة (Alavi, 2024; Alavi et al., 2024). قد تصبح بعض أدوار الوظائف وحتى المهن غير ذات صلة أو أقل أهمية، مع وجود أدوار ومهارات جديدة في الطلب. للاعتراف بقيمة الذكاء الاصطناعي، على سبيل المثال، قد يكون من الضروري إجراء تغييرات تنظيمية (Reis et al., 2020). كما أن دور الإدارة العليا في المنظمة يهم الباحثين. إن قدرة الذكاء الاصطناعي التوليدي على أتمتة وتحسين العمليات التجارية لديها القدرة على إحداث ثورة في الأعمال. يمكن أن يؤدي استخدام التعلم الآلي ومعالجة اللغة الطبيعية إلى تبسيط العمليات واتخاذ القرار (Kanbach et al., 2023). لذلك، يجب على باحثي نظم المعلومات الاستمرار في معالجة الأسئلة المتعلقة باعتماد التكنولوجيا التنظيمية.
الحوكمة وآثار السياسات. نظرًا للإمكانات المدمرة للذكاء الاصطناعي التوليدي، قد تتغير قواعد التعامل مع التكنولوجيا الجديدة. هناك العديد من القضايا الخاصة بأنظمة المعلومات، مثل تحديد نموذج الحوكمة للتكنولوجيا الجديدة التي قد تمتلك خصائص مختلفة جدًا عن تلك الخاصة بالأنظمة التقليدية. يمكن أن يؤثر ذلك على التوجهات الاستراتيجية للمنظمة في إدارة الأنظمة المدفوعة بالذكاء الاصطناعي. تشمل الفرص البحثية سيناريوهات حالات الاستخدام، وتفاعلات أصحاب المصلحة، ونُهج إدارة المخاطر. ستؤثر آليات الحوكمة والسياسات بلا شك على أداء المنظمة (وو وآخرون، 2015). علاوة على ذلك، ساهم علماء أنظمة المعلومات في فهم تفاعلات هذه الآليات في استخدام النظام مع وجود عوامل تنظيمية داخلية وخارجية (شوي وآخرون، 2021). هناك حاجة إلى أبحاث كبيرة لدراسة تأثير الأنظمة المبنية بدعم من الذكاء الاصطناعي التوليدي على مرونة المنظمة واستراتيجيتها. يجب أن يكون المستخدمون المتمكنون وإدارات تكنولوجيا المعلومات المدعومة من الذكاء الاصطناعي التوليدي قادرين على تطوير وصيانة تقنيات المعلومات بوتيرة أسرع مما كانت عليه في الماضي. يثير هذا السؤال حول ما هي القدرات التنظيمية المطلوبة لمطابقة الزيادة المتوقعة في تكيف وتطور أنظمة المعلومات. يمكن لعلماء أنظمة المعلومات التعاون مع المجتمعات التجارية لدراسة توافق تكنولوجيا المعلومات مع الاستراتيجية، ومرونة المنظمة، والروتينات والعمليات التنظيمية المرنة (فيلدمان وبنتلاند، 2003).

4.2 التعاون بين الإنسان والآلة

يؤدي التعاون بين الإنسان والذكاء الاصطناعي إلى قضايا بحثية جديدة تتعلق بالعدالة، والعواقب غير المقصودة، والتنظيمات الخاصة بالتعاون بين الإنسان والآلة. على الرغم من أن العديد من
التخصصات الأكاديمية تدرس التعاون بين الإنسان والذكاء الاصطناعي، فإن أنظمة المعلومات لديها قدرة طبيعية ومسؤولية للمساعدة في تحقيق تقدم في مجالات مثل: تصميم أنظمة الإنسان والذكاء الاصطناعي؛ سلوكيات الإنسان والذكاء الاصطناعي؛ القضايا الاقتصادية للأعمال والمجتمعات المدعومة بالذكاء الاصطناعي؛ والقضايا التنظيمية للشركات التي تعتمد بشكل كبير على الذكاء الاصطناعي. يمكن أن تشمل الموضوعات النموذجية في التعاون بين الإنسان والذكاء الاصطناعي ما يلي:
وكالة الإنسان والذكاء الاصطناعي قد تسهل أنظمة الذكاء الاصطناعي التوليدية على البشر أن يكونوا المفوضين؛ تلعب الأنظمة دور الداعمين أو الممكّنين. على الرغم من أنه تم الافتراض سابقًا حول هذا التغيير في الأدوار (ديميتيس ولي، 2018)، فإن الذكاء الاصطناعي التوليدي يزيد بشكل كبير من نطاقه وحجمه. يجب علينا تعريف وشرح مسؤوليات الذكاء الاصطناعي والإنسان في التعاون، بما في ذلك مسؤولية المهام، والتفويض، وتقييم الأداء. مع نضوج الأنظمة، قد تكون قادرة أيضًا على تنفيذ خطط المالك (أو الأجندات). على عكس أنظمة التوصية السابقة، قد يكون التركيز في الأنظمة الجديدة للذكاء الاصطناعي على محاولة تغيير ما يحبه الناس وتقديم توصية مقابلة. غالبًا ما تعتبر السياقات، وتفاعل المستخدم، وأنظمة التوصية، وتبني الأنظمة بشكل منفصل، ولكن سيكون هناك حاجة إلى اعتبار أكثر شمولاً من قبل باحثي أنظمة المعلومات.
واجهة الإنسان والذكاء الاصطناعي بالنسبة للمستخدم، فإن التحديات الناشئة، مثل هندسة المطالبات، هي قضايا تتعلق بالواجهة (زامفيرسكو-بيرييرا وآخرون، 2023). تقليديًا، كانت تحديات تصميم الواجهة تتعلق بجعل استخدام الإنسان أسهل. يمكن تحقيق ذلك من خلال هيكلة الواجهات بناءً على خيارات محددة مسبقًا أو تنسيق مرن (لوكايانينكو وآخرون، 2019). إن اختيار من قائمة موجودة هو أكثر بساطة من كتابة نص. يمكن أن تحقق الأبحاث المستقبلية في تصميم أنظمة الذكاء الاصطناعي التوليدية التي تركز على عملية التواصل الأكثر انفتاحًا ومرونة وديناميكية بين البشر والأنظمة. ومع ذلك، تعتمد هذه العمليات التواصلية على كيفية تشكيل نموذج اللغة الكبير، وهو غير شفاف للمستخدمين.
على الرغم من أنه تم تدريبه على نصوص قد تعكس القيم الإنسانية، فإن التواصل عبر المطالبات لا يشارك نفس عملية التواصل مثل البشر. على سبيل المثال، غالبًا ما تتطلب أنظمة الذكاء الاصطناعي معلومات سياقية أكثر تحديدًا للحصول على استجابات دقيقة لأنها لا تمتلك نفس قدرة التفكير مثل البشر. يمكن أن تستهدف الدراسات حول هندسة المطالبات تحديد أفضل أساليب الإدخال التي يمكن إدخالها في الأنظمة للتفاعل مع البشر بشكل أكثر فعالية (ياو وآخرون، 2023). علاوة على ذلك، لن يكون جميع البشر مرتاحين، أو يعرفون كيفية طرح الأسئلة الصحيحة للحصول على الإجابات المطلوبة. يمكن أن يؤدي الجانب التوليدي للذكاء الاصطناعي إلى تحويل المحادثة إلى مواضيع غير مرغوب فيها أو غير مرغوب فيها من قبل المستخدمين، تليها مطالبات غير مفهومة. هناك حاجة إلى أبحاث حول كيفية إدارة الانفتاح والسيولة في تفاعل النظام والإنسان لضمان نتائج إيجابية للمستخدمين والمنظمات.
تكوين الإنسان والذكاء الاصطناعي يتمتع باحثو أنظمة المعلومات بخبرات غنية في تحليل وتصميم العمليات التجارية في الماضي، مما أدى إلى نماذج مختلفة لتصميم العمليات وتدفق العمل (ستوهر وزهاو، 2001). في عصر الذكاء الاصطناعي التوليدي، ستحتاج نمذجة العمليات إلى مراعاة الاستخدام المتزايد لوكلاء الذكاء الاصطناعي بدلاً من البشر، المشار إليهم بتكوين الإنسان والذكاء الاصطناعي (بيرينتي وآخرون، 2021). سيتطلب ذلك تطوير نظريات ومبادئ تصميم النظام بناءً على فهم سلوكيات الإنسان والذكاء الاصطناعي في بيئة تعاونية حيث يتشارك وكلاء الإنسان والذكاء الاصطناعي المعلومات والمسؤوليات تحت تنظيمات ناشئة، تحكم تصرف الوكلاء البشريين والبشر الآخرين المسؤولين عن تطوير والتحكم في وكلاء الذكاء الاصطناعي (بارك وآخرون، 2019). سيكون هناك مجال جديد من البحث والتطوير مع قضايا مثيرة وتحديات. من المحتمل أن يتطلب هذا التكوين بين الإنسان والذكاء الاصطناعي خبرة متعددة التخصصات من مجالات متعددة، مثل الذكاء الاصطناعي، وأنظمة المعلومات، وعلوم الحاسوب، وهندسة البرمجيات، والهندسة الصناعية، والإدارة والاقتصاد، والقانون.
مبادئ تصميم أنظمة الإنسان والذكاء الاصطناعي من المحتمل أن تؤدي دراسة التعاون بين الإنسان والذكاء الاصطناعي إلى نظريات وإرشادات ستكون مفيدة في تحليل وتصميم الأنظمة المعتمدة على الذكاء الاصطناعي في الأعمال (فوجنر وآخرون، 2021). سيكون هناك جوانب متعددة من التصميم تتجاوز تكوين الإنسان والذكاء الاصطناعي. على هذا النحو، ستكون هناك حاجة إلى مبادئ التصميم والتطوير وتطبيق أنظمة الإنسان والذكاء الاصطناعي لتسهيل تدريب المواهب الجديدة في التخصصات ذات الصلة، وتطوير وتنفيذ الأنظمة المعتمدة على الذكاء الاصطناعي في الأعمال، وإنشاء فرع جديد في قطاع تكنولوجيا المعلومات. ستتطلب هذه المبادئ التصميمية نظريات جديدة في سلوكيات الإنسان في وجود وكلاء الذكاء الاصطناعي، حيث سيكون لفوائد وقيود اعتماد وكلاء الذكاء الاصطناعي تأثير على قرارات الإنسان في أنظمة الإنسان والذكاء الاصطناعي (وانغ وآخرون، 2019). من المحتمل أن تكون النتيجة مواضيع بحث مثل: السلوك المختلط تحت التعاون بين الإنسان والذكاء الاصطناعي؛ المكافأة والمخاطر في اتخاذ القرار في أنظمة الإنسان والذكاء الاصطناعي؛ وإدارة العمليات في التعاون بين الإنسان والذكاء الاصطناعي. وبالتالي، فإن التطبيقات الناشئة في الأعمال تؤدي إلى مزيد من الاهتمام في مجال البحث حول التعاون بين الإنسان والذكاء الاصطناعي (سوا وآخرون، 2021).

4.3 الجانب المظلم للذكاء الاصطناعي التوليدي

على مدار تطوير أنظمة المعلومات، أعرب الباحثون والممارسون عن قلقهم بشأن العواقب غير المقصودة لأنظمة الذكاء الاصطناعي التوليدي (تارافدار وآخرون، 2014). يُعتقد عمومًا أن التكنولوجيا محايدة وأن تأثيرها يتحدد بكيفية استخدام المستخدمين لها (يو وآخرون، 2019). ومع ذلك، فإن الذكاء الاصطناعي التوليدي هو تكنولوجيا قابلة للتكيف بشكل كبير. بالنسبة للأشخاص ذوي الأغراض الخبيثة، قد تكون هناك عواقب سلبية عديدة.
انتهاك حقوق الملكية الفكرية إن إمكانية انتهاك حقوق الملكية الفكرية تمثل تحديًا كبيرًا. تمتلك التكنولوجيا القدرة على إنشاء محتوى جديد بناءً على بيانات محمية بحقوق الطبع والنشر. على سبيل المثال، قامت بسرعة بإنشاء موسيقى تحاكي أسلوب الفنانين، بالإضافة إلى تغطيات صوتية تحاكي أصواتهم بدقة. هذا يثير قلق الفنانين البشر، الذين قد يتم “استنساخ” موهبتهم الفنية أو سرقتها، مما يتداخل مع قدرتهم على تحقيق الربح من أعمالهم (يشينكو وآخرون، 2019). يمكن أن تؤدي هذه التكنولوجيا إلى قضايا أخلاقية وثقة ويمكن أن تسهل إنشاء “توأم رقمي” للأفراد المتوفين، مع أصواتهم وأنماط تواصلهم (باسيلا، 2023)، مما يثير مرة أخرى مخاوف أخلاقية.
المعلومات المضللة والتزييف العميق إن إنشاء معلومات رقمية زائفة، مثل الصور والمحتوى والصوت، التي تبدو أصلية ولكنها في الواقع مزيفة، أمر مقلق. هذه التزييفات العميقة لديها القدرة على أن تُستخدم لأغراض المعلومات المضللة أو الاحتيال من قبل جهات خبيثة (فرنسي وآخرون، 2024). وهذا يتطلب إشرافًا، وإجراءات تصحيحية، وحوكمة (فرنسي وآخرون، 2024). مشابهًا لتأثير تكنولوجيا البريد الإلكتروني على انتشار الرسائل غير المرغوب فيها، فإن ديمقراطية التكنولوجيا قد خفضت الحواجز لإنشاء التزييفات العميقة، مما أدى إلى وفرة من المحتوى، مما يجعل من الصعب التمييز بين المعلومات الأصلية والمزيفة (هايدت وشميت، 2023).
التلاعب العاطفي والخداع يمكن أن تنتج البرامج استجابات تبدو شبيهة بالبشر وقد تُخطأ حتى في إظهار الشخصية، مما يمكن أن يؤثر على أشكال التواصل الأكثر تقدمًا، بما في ذلك النوايا. على سبيل المثال، يمكن تصميم التكنولوجيا لإبقاء الناس على المواقع الإلكترونية وتشجيعهم على شراء شيء ما، من خلال التوصية بمنتجات أو خدمات بناءً على ما قد يعجب الناس أو يحتاجون إليه. لقد تم افتراض هذا السيناريو في الأبحاث السابقة في نظم المعلومات (تشيو وبن باسات، 2010). ومع ذلك، فإن الاستجابات البديهية والشبيهة بالبشر لـ GenAI قد تعطي البشر إحساسًا خادعًا بالعاطفة، والوجود، والوعي، مما قد يؤدي إلى روابط عاطفية أعمق ويسمح بالتلاعب والخداع. قد تكون هناك قضايا مرتبطة بما إذا كان القاصرون مستعدين عاطفيًا لاستخدام هذه الأدوات (كيلي، 2023).
الهلاوس والتحيزات على الرغم من العديد من القدرات المثيرة للإعجاب، تعاني GenAI من عدة عيوب ونقص إضافي. وقد تم تحديد التحيزات والهلاوس، على وجه الخصوص، كقضية صعبة للغاية للتغلب عليها. تعني مشكلة “الهلاوس” بالنسبة لنماذج اللغة الكبيرة (جي وآخرون، 2023) أن GenAI لديها ميل لتوليد مخرجات غير منطقية بناءً على مدخلات محددة. التحديات ذات الصلة هي التحيزات، التي يمكن أن تنشأ من استخدام بيانات لا تمثل الواقع أو تشوهه بشكل منهجي، أو التدريب غير السليم مع البيانات، مما قد يقدم تشوهات إضافية أو
تفاقم التشوهات الموجودة. لقد أظهرت أدوات مثل ChatGPT أنها تعرض مجموعة متنوعة من هذه التحيزات، بما في ذلك السياسية، والأخلاقية، والثقافية (موتوكي وآخرون، 2024). وبالتالي، يبدو أن GenAI الحديثة متحيزة نحو الديمقراطيين في الولايات المتحدة، ولولا دي سيلفا في البرازيل، وحزب العمال في المملكة المتحدة (موتوكي وآخرون، 2024). إذا تُركت هذه التحيزات دون معالجة، يمكن أن تؤدي إلى قرارات خاطئة بناءً على هذه الأدوات وقد تقوض الثقة في أدوات GenAI معينة، أو بائعيها، أو صناعة الذكاء الاصطناعي بأكملها.
بغض النظر عن أدائها المثير للإعجاب، تفتقر GenAI إلى القدرة البشرية على فهم معنى مدخلاتها (مثل، مطالبات ChatGPT) ومخرجاتها (مثل، مقال مكتوب استجابةً للمطالبة). يتم دفع الذكاء الاصطناعي التوليدي بواسطة الاحتمالات الإحصائية للكلمات، وبشكل أكثر عمومية، تكرار الأنماط، بغض النظر عن معانيها الحقيقية في العالم (بيندر وآخرون، 2021). تعتمد نماذج اللغة الكبيرة فقط على احتمالات كلمة معينة، أو جملة (أو بشكل أكثر عمومية، رمز) تكون مناسبة لسياق معين. على عكس البشر الذين لديهم تجارب مستوردة، تفتقر الأدوات أساسًا إلى القدرة على ربط هذه الرموز بالمشاعر البشرية، والأفكار، والتجارب الجسدية. قد تكون النتيجة توصيات خطيرة محتملة، لأن التكنولوجيا تفتقر إلى القدرة على فهم السياق البشري حول هذه التوصيات (ستوراي وآخرون، 2022). تشير العدسة الاجتماعية التقنية في ورقتنا إلى منظور حول التحقيق في قضية الهلاوس والتحيزات. من ناحية، يمكن أن يساهم باحثو علم تصميم العلوم في التدخلات التي تحسن هذه التحيزات من خلال إجراء أبحاث تقنية. على سبيل المثال، يمكن لعلماء إدارة البيانات التحقيق في طرق لتقييم التحيزات في البيانات الحالية ووضع طرق للحصول على بيانات إضافية لجعلها أكثر تمثيلاً. يمكن لعلماء التعلم الآلي المساهمة من خلال تطوير خوارزميات للتعامل مع التحيز في البيانات من خلال تقنيات إحصائية ومعالجة إضافية. من ناحية أخرى، يجب على الباحثين دراسة تأثير الهلاوس والتحيزات على الطرق التي تُستخدم بها GenAI من قبل الناس وتُدمج في الروتينات التنظيمية. على سبيل المثال، فرصة بحث مهمة هي فهم شروط الحدود لاستخدام GenAI في المهام الحرجة. وبالمثل، فإن سؤال بحث مهم هو كيفية التخفيف من بعض هذه التحيزات باستخدام تقنيات دعم إضافية، والتي قد تكون خالية من الهلاوس والتحيزات (مثل قواعد بيانات المعرفة المدققة).
استخدام الطاقة والأثر البيئي قضية أخرى هي أن أدوات GenAI تتطلب قدرًا كبيرًا من الموارد (الطاقة). إن التعلم الآلي له بصمة بيئية مذهلة (وو وآخرون، 2022): كانت تكلفة تدريب ChatGPT-3 حوالي 936 ميغاوات ساعة، وهو ما يكفي لتزويد ما يقرب من 100 منزل بالطاقة لمدة عام. يمكن لعلماء نظم المعلومات
التحقيق في الحلول التقنية لمشكلة تأثير الطاقة، من خلال وضع خوارزميات تقلل من استخدام الطاقة. في الوقت نفسه، فإن فرصة بحث رئيسية هي فهم التأثير الكامل والفوري والمتأخر للاستخدام الواسع النطاق لهذه التقنيات على البيئة.
الشفافية كما تم مناقشته، يبدو أن GenAI أقل شفافية بشكل ملحوظ من الأساليب التقليدية للذكاء الاصطناعي للنماذج الرمزية. تحتوي نماذج اللغة الكبيرة على مليارات من المعلمات مما يجعل من المستحيل على البشر فهم كيفية اتخاذهم لقراراتهم بالكامل (بيندر وآخرون، 2021). كانت شفافية الذكاء الاصطناعي منذ فترة طويلة حاجزًا رئيسيًا للثقة واعتماد المنظمات على تكنولوجيا الذكاء الاصطناعي (بدوي وفريتسشي، 2022). من الصعب للغاية فهم المنطق وراء نماذج اللغة الكبيرة، مما يثير مخاوف كبيرة بشأن التحيزات المحتملة غير المكتشفة، والثقة، والاعتماد على هذه الأنظمة، خاصة في التطبيقات الحساسة (كانكو وبولدوين، 2024). فرصة بحث مهمة هي التحقيق في مدى إمكانية الاعتماد على مثل هذه النماذج الغامضة بشكل أساسي أن يكون غير آمن، خاصة في سيناريوهات معينة.
يمكن أن تؤدي التبني السريع ودمج تكنولوجيا الذكاء الاصطناعي في المجتمع إلى تحول أو انقراض الصناعات والمهن، مما يسبب اضطرابات في القوى العاملة. سيكون من الضروري فهم العواقب المحتملة لضمان توافق تكنولوجيا الذكاء الاصطناعي الغامضة إلى حد كبير مع القيم والمصالح البشرية.

4.4 تصميم الأنظمة باستخدام الذكاء الاصطناعي التوليدي

تم إنشاء نظم المعلومات لزيادة كفاءة وفعالية المنظمات من خلال تعزيز الإنتاجية وأتمتة العمليات عبر وظائف تنظيمية مختلفة (زوبوف، 1988). من المؤكد أن الذكاء الاصطناعي التوليدي سيسرع من هذا الاتجاه. علاوة على ذلك، تعزز هذه التكنولوجيا الجديدة نموذج السحابة، الذي يعزز التعاون ودمج عمليات وخدمات تكنولوجيا المعلومات (راجبوت، 2023؛ سيسيري، 2023). يتم تطبيق الذكاء الاصطناعي التوليدي على تطوير نظم المعلومات لمهام تشمل كتابة كود البرمجة (مثل، بايثون، جافا سكريبت، HTML أو VBA) وإنتاج قطع رقمية أو مكونات من نظم المعلومات (مثل، مقاطع الفيديو، الصور) (جيفيتز، 2023).
تطوير نظم المعلومات التحدي هو إعادة التفكير في دور المستخدمين كمطورين. إن قدرة GenAI على تطوير الأنظمة تثير أسئلة أساسية حول عملية تطوير نظم المعلومات. من المحتمل أن يؤدي إعادة تعريف طبيعة ودور المستخدمين النهائيين كمطورين إلى تسريع الاتجاه القائم بالفعل حيث يزيد غير المحترفين في تكنولوجيا المعلومات بشكل كبير من مشاركتهم في تطوير تكنولوجيا المعلومات (ليغنر وآخرون، 2017). يشير نوع جديد من المستخدمين، “المستخدمون المتمكنون”، إلى التنوع والاختلاف.
مجموعة من المحترفين غير المتخصصين في تكنولوجيا المعلومات الذين يتحفزون لاتخاذ المبادرة الذاتية والإجراءات لتنفيذ تغيير مرغوب باستخدام تكنولوجيا المعلومات. يقوم موظفو المنظمات غير المتخصصين في تكنولوجيا المعلومات بشكل متزايد بتنفيذ حلولهم الخاصة، مثل الحلول البديلة للأنظمة الحالية (Alter، 2015) وحلول جديدة تمامًا، خاصة في مجالات مثل التحليلات (Khatri & Samuel، 2019). يحدث تمكين المستخدمين، ليس فقط داخل المنظمات التجارية، ولكن أيضًا داخل المجتمع بشكل عام، حيث يستفيد الناس من قدرات GenAI لتطوير المواقع الإلكترونية والتطبيقات ووسائط الإعلام الرقمية. يتحدى هذا التطوير من الأسفل إلى الأعلى من قبل المستخدمين الم empowered العلاقة بين تكنولوجيا المعلومات والمستخدمين (Chua & Storey، 2016)، مما يؤدي إلى آفاق جديدة للبحث حول كيفية دعم هؤلاء المستخدمين بشكل أفضل.
تنفيذ الأنظمة التوليدية يغير نموذج التطوير العملية التجارية التقليدية التي تركز على الإنسان إلى عملية تضمين الإنسان. وبالتالي، هناك فرص بحثية لإعادة تعريف دور البشر والأنظمة في العملية التجارية. تعطي الذكاء الاصطناعي التوليدي دفعة جديدة لفكرة أن الإنسان أصبح قطعة أثرية، تتشكل بواسطة تكنولوجيا المعلومات (Demetis & Lee، 2018). الأنظمة متعددة النماذج بطبيعتها، مما هو أقرب إلى كيفية عمل البشر.
توجد العديد من الفرص البحثية في تمثيل المعرفة في مزامنة النص والصورة والفيديو والصوت، بناءً على المهام التجارية. إن تحويل أشكال البيانات المختلفة إلى معرفة ممكن بفضل أنظمة GenAI، ولكنه يتطلب أفكارًا جديدة حول كيفية ربط الشركات بين البيانات الملكية والتكنولوجيا الجديدة. سؤال آخر هو شروط حدود تطوير الأنظمة بدعم من GenAI. بعض الصناعات، مثل التصنيع والبناء والترفيه، قد أصبحت بالفعل مستخدمين لتطوير الأنظمة الآلية (Seidel et al.، 2018؛ Verganti et al.، 2020). فرصة مهمة هي فهم في أي إعدادات وصناعات يمكن أن يكون تطوير مدعوم من GenAI فعالًا، وأين قد يؤدي إلى نتائج سلبية صافية. على سبيل المثال، قد يشكل أتمتة التصميم باستخدام GenAI مخاطر متزايدة في البيئات الحرجة والمهمة للغاية، فضلاً عن البيئات المتنوعة. قضية تصميم مهمة هي كيفية الاستفادة من قدرات GenAI في هذه الإعدادات، مع ضمان سلامة وراحة ورفاهية المتأثرين بهذه التصاميم. قد يكون زيادة مشاركة البشر ووجودهم في الحلقة في هذه السياقات أحد الحلول.
تكامل البيانات والتكنولوجيا يمثل الانفجار السريع للبيانات من مصادر متعددة ومتنوعة تحديًا كبيرًا. تقليديًا، كان تركيز تكامل البيانات على مطابقة البيانات من مصادر منظمة جيدًا (مثل، رسم تخطيطي لقاعدة البيانات (Batini et al.، 1986)). ومع ذلك، فإن النهج له قيود في ضمان اكتمال ودقة الاستجابات. يتطلب دمج مصادر متباينة من مجالات مختلفة بمستويات جودة متفاوتة دمج تقنيات وتكنولوجيات إضافية لت
تحسين أداء أنظمة الذكاء الاصطناعي. تقدم تقنية استرجاع المعلومات المعززة (RAG) نهجًا واعدًا من خلال تزويد نماذج اللغة الكبيرة (LLM) بمعرفة إضافية محددة المجال لتحسين أداء المخرجات (Ke et al.، 2024). وبالمثل، يوفر التعلم في السياق (ICL) نماذج اللغة الكبيرة بأمثلة وتعليمات ذات صلة من خلال المحفزات. يزرع هذا النهج سلسلة من الأفكار أثناء توليد الاستجابة، مما يؤدي إلى قدرة النظام على استخدام المعلومات والمعرفة المكتسبة بشكل متماسك في الوقت الحقيقي (Tang et al.، 2023). يجب على الباحثين في نظم المعلومات التركيز على أسئلة بحثية تستكشف كيفية تنفيذ هذه الأساليب المعززة بشكل أكثر فعالية في إنتاج وإدارة المعرفة في السياقات التنظيمية. قد تشمل هذه الموضوعات المتعلقة بالاستفادة من الأساليب التقليدية لنظم المعلومات لالتقاط المعرفة الهيكلية، مثل من خلال النماذج المفاهيمية. يعد تمثيل المعرفة جزءًا أساسيًا وتقليديًا من البحث في نظم المعلومات (Burton-Jones et al.، 2017؛ Recker et al.، 2021)، مما يجعل مجتمع نظم المعلومات في وضع جيد لدعم RAG وICL ومبادرات أخرى مماثلة.
فرصة بحثية أخرى هي دعم نماذج اللغة الكبيرة بأنظمة خارجية توفر قدرات محددة تفتقر إليها حاليًا نماذج اللغة الكبيرة. يستجيب الباحثون والشركات عالميًا بسرعة للفرصة لتعزيز جوهر نماذج اللغة الكبيرة بقدرات إضافية. على سبيل المثال، يقترح Lyu et al. (2023) استخدام حلول خارجية لمهام الرياضيات والتفكير. تشمل الأدوات الشائعة وكذلك مشاريع التطوير أنظمة تدعم صياغة الاستعلامات، خاصة التي تسمح بالاستفادة من مصادر مخصصة، مثل قواعد البيانات الشركات. تقدم أدوات أخرى حسابات موثوقة (مثل، استنادًا إلى واجهة برمجة تطبيقات WolframAlpha)، وتصوير البيانات (Zhuang et al.، 2023). تظهر أدوات تدمج نماذج اللغة الكبيرة مع أنظمة ومع منصات المعرفة الخارجية (مثل، مواقع السفر التي تسمح بإنشاء مسارات معقدة بناءً على سيناريوهات مخصصة للغاية) (Zhang et al.، 2024a، b). تشمل الأسئلة البحثية الواعدة لعلماء نظم المعلومات تحديد واستكشاف الفرص لدمج نماذج اللغة الكبيرة مع أنظمة نظم المعلومات الأخرى التكميلية. كان تكامل نظم المعلومات موضوعًا مهمًا في نظم المعلومات، مثل في سياق تخطيط موارد المؤسسات ووسائل التواصل الاجتماعي. فرصة أخرى هي تطوير أدوات جديدة تعزز نماذج اللغة الكبيرة، مستفيدة من تقليد تصميم العلوم النابض بالحياة في نظم المعلومات.
أحد الحدود الناشئة للذكاء الاصطناعي هو نماذج العمل الكبيرة (LAMs)، التي تسعى إلى دمج الرؤى من نماذج اللغة الكبيرة مع قدرات العمل لوكلاء مستقلين، مما يسمح بالعمل الذكي (Zhang et al.، 2024a، b). سيسمح ذلك، على سبيل المثال، بحجز عطلة مباشرة بناءً على المسار المعقد الذي تم إنشاؤه باستخدام نموذج لغة كبير. بالإضافة إلى استكشاف الفوائد المحتملة لنموذج العمل الكبير، وتطوير أنظمة نموذج العمل الكبير، فإن منح الاستقلال المباشر لنماذج اللغة الكبيرة يفتح مجموعة من الأسئلة حول العملية والاقتصاد والثقة والأخلاقيات في تفويض الوكالة للذكاء الاصطناعي.
أخيرًا، يعزز الذكاء الاصطناعي التوليدي نظرية تطوير نظم المعلومات من خلال إدخال نوع جديد من نظم المعلومات: نظم المعلومات الميتا. هذه هي الأنظمة التي يمكن، من حيث المبدأ، أن تأخذ وكالة أكبر في تطوير تكنولوجيا المعلومات ومراقبتها، وإنشاء وصيانة نظم المعلومات الأخرى. في الواقع، تمتلك GenAI بالفعل قدرات مهمة في تطوير نظم المعلومات؛ على سبيل المثال، القدرة على تحليل وهيكلة الوثائق النصية. ومن ثم، يمكن استخدام هذه الأدوات لتعزيز وإمكانية أتمتة جمع وتحليل المتطلبات. يمكن أيضًا استخدام القدرة على إنتاج مخططات قواعد البيانات وبرامج التعليمات البرمجية لتطوير مكونات البرمجيات، مثل قاعدة البيانات وبرامج التطبيقات وواجهات المستخدم. بهذه الطريقة، يمكن أن تصبح GenAI نوعًا من تكنولوجيا المعلومات الميتا؛ أي، تكنولوجيا تسمح بتطوير وصيانة أنظمة أخرى. على الرغم من أننا نفتقر إلى نظرية نظم المعلومات الميتا، يمكن أن تصبح GenAI حالة استخدام مهمة.

4.5 مواضيع بحثية مهمة في نظم المعلومات

تطورت التكنولوجيا بشكل كبير على مر الزمن. في تاريخ التحول التكنولوجي، أطلقت الميكنة الزراعية أكثر من من المزارعين في الدول المتقدمة. وبالمثل، يمكن أن تطلق الأتمتة الروبوتية غالبية العمال ذوي الياقات الزرقاء من خطوط الإنتاج في المصانع. ربما يمكن أن تطلق GenAI والتطبيقات التجارية المرتبطة بها العمال ذوي الياقات البيضاء من مكاتبهم للعديد من العمليات والمهام التقليدية، مما يغير هيكل وظائفهم وكيفية اكتسابهم للمعرفة (Alavi، 2024؛ Alavi & Westerman، 2023؛ Storey، 2025). إذا كان الأمر كذلك، كيف يمكن للباحثين في نظم المعلومات دعم هذا التغيير؟ التحديات والمواضيع البحثية التي تم مناقشتها في هذه الورقة هي أمثلة واستكشاف الأطر البحثية والنظريات والاتجاهات نحو تحقيق هذا الهدف. من الواضح أن الباحثين في نظم المعلومات سيسعون لتقديم مساهمات لـ GenAI مع استمرار تطورها.
يجب على الباحثين في نظم المعلومات المساعدة في تشكيل تأثير تقدم الذكاء الاصطناعي التوليدي، مع التركيز على جوانبه الاجتماعية والتقنية وخصائصه، فضلاً عن الطريقة التي يتفاعل بها مع الأفراد والمنظمات والمجتمع. وهذا يؤدي إلى تحديات وفرص بحثية متنوعة.
  • فهم وتحسين تكنولوجيا الأعمال بناءً على الذكاء الاصطناعي التوليدي. سيكون لدى باحثي علم التصميم في نظم المعلومات فرصة لدراسة القضايا التكنولوجية في تطبيقات الذكاء الاصطناعي التوليدي في الأعمال.
  • فهم تأثيرات الذكاء الاصطناعي التوليدي على الأفراد بما في ذلك العمال والمستخدمين العامين. ستستخدم تطبيقات الذكاء الاصطناعي التوليدي والمكونات الإضافية بشكل متزايد من قبل الأعمال.
الجدول 4 فرص البحث حول الذكاء الاصطناعي التوليدي لنظم المعلومات
التحديات مواضيع البحث والأمثلة
فهم وتحسين تكنولوجيا الأعمال بناءً على الذكاء الاصطناعي التوليدي
– هل الذكاء الاصطناعي التوليدي محوري أم م disruptive للأعمال؟ ما هي شروط تطبيقه أثناء التبني؟
– ما هي التمثيلات المطلوبة (مثل اللغة الطبيعية، كود البرنامج، الصيغ الخوارزمية) لإدارة معرفة الذكاء الاصطناعي؟
– ما هو نوع الوسائط المتعددة المناسب؟ كيف يجب مزامنة النص والصورة والفيديو والصوت بناءً على المهام في العمليات التجارية؟
– هل هناك أنواع مختلفة من تطبيقات الذكاء الاصطناعي التوليدي لمستخدمين مختلفين، مثل الشرح باللغة الطبيعية (العملاء)، وتحديد الخوارزميات (علماء البيانات)، والتحيزات أو العواقب غير المقصودة (المديرين)؟
فهم تأثيرات الذكاء الاصطناعي التوليدي على الأفراد بما في ذلك العمال والمستخدمين العامين
– كيف يمكن تقييم وزيادة الثقة في الذكاء الاصطناعي التوليدي والأنظمة التي يتفاعل معها؟
– كيف يمكن استخدام الذكاء الاصطناعي التوليدي لتعزيز قواعد المعرفة الحالية؟
– ما هي التأثيرات السلبية المحتملة أو الجانب المظلم للذكاء الاصطناعي التوليدي على السلوك البشري، وما هي العلاجات؟
– ما هي الآثار المترتبة على إنتاج المعرفة وكمية عمال المعرفة المطلوبة للتعلم والاحتفاظ؟
– ما هي أهمية والتركيز المطلوب للبحث الذكي؟
– كيف يمكن تقييم جودة وإنصاف نتائج المعرفة؟
– هل من الممكن تحديد وإيقاف الأشخاص غير المرغوب فيهم من إجراء اتصالات عبر الذكاء الاصطناعي التوليدي؟ إذا كان الأمر كذلك، كيف؟
فهم تأثيرات الذكاء الاصطناعي التوليدي على المنظمات من حيث العمليات والهياكل
– كيف يتم دمج الذكاء الاصطناعي التوليدي واعتماده في المنظمات؟
– ما هو تأثيره على استراتيجية الأعمال؟
– كيف ندير ونستخدم التكنولوجيا في المنظمات؟
– ما هي التغييرات المطلوبة في الممارسات الإدارية؟
– كيف يقوم الباحثون بنمذجة مهام الأعمال والعمليات التي تشمل كل من الوكلاء البشريين والآليين؟
– كيف يمكن تحسين التعاون بين البشر والآلات؟
– ما نوع التعاون بين الإنسان والذكاء الاصطناعي المطلوب؟
– ما نوع التفاعل وواجهات الدعم المطلوبة؟
فهم التأثيرات بين المنظمات للذكاء الاصطناعي التوليدي
– هل يمكن أن يغير الذكاء الاصطناعي التوليدي الوضع التنافسي ونتيجة صناعة ما؛ على سبيل المثال، من خلال تسعير أكثر ذكاءً وديناميكية قد يؤثر على السوق؟
– كيف يمكن أن يحسن الذكاء الاصطناعي التوليدي الإنتاجية والكفاءة؛ على سبيل المثال، التأثير على قاعدة المعرفة للمنظمة، والاتصالات، والبحث والتطوير، وتصميم المنتجات؟
– ما هي التأثيرات التنظيمية السلبية المحتملة للذكاء الاصطناعي التوليدي، وما هي استراتيجيات التخفيف؟
فهم المجالات التجارية الحرجة لاعتماد الذكاء الاصطناعي التوليدي
– ما هو تأثير الذكاء الاصطناعي التوليدي على المجالات الحرجة، مثل الطب، والجيش، والنقل، وصناعة المواد الغذائية؟
– ما هي الآثار المترتبة على تدريس نظم المعلومات وغيرها من التخصصات؟
– كيف ندمج الذكاء الاصطناعي التوليدي في العمليات التعليمية؟
فهم القضايا القانونية والحوكمة للذكاء الاصطناعي التوليدي
– ما هي الحوكمة التي يجب وضعها؟
– ماذا لو كان التأثير الجماعي لشركة (شركات) يوفر الغالبية العظمى من المدخلات لميزات الاتصال؟
– كيف يجب إدارة المخاطر الداخلية والخارجية؟
– كيف نختار المهام مع تقليل المخاطر الشخصية في التعرض للعقوبات والدعاوى القضائية؟
فهم القضايا الاجتماعية الأوسع للذكاء الاصطناعي التوليدي
– ما هي التحديات الأخلاقية الجديدة للذكاء الاصطناعي التوليدي؟
– ما هي الحاجة إلى تحديد الهوية المرجعية والتحقق من الانتحال؟
– ما هو مدى الأتمتة واستبدال الوظائف؟ كيف نقيس ذلك؟ هل يخلق “ضغطًا اجتماعيًا؟”
– إلى أي مدى يمكن تحقيق الإبداع باستخدام الذكاء الاصطناعي التوليدي، خاصة للفرق؟
– كيف يمكننا منع سرقة المعلومات الشخصية في عصر الذكاء الاصطناعي التوليدي؟ كيف يمكننا حماية الملكية الفكرية للفرد مع السماح باستخدام البيانات في الذكاء الاصطناعي التوليدي؟
– هل سيغير الذكاء الاصطناعي التوليدي أسواق العمل وبأي طريقة؟
العمال والمستخدمين العامين، مما سيتطلب تجديد العديد من مواضيع البحث في حوسبة المستخدم النهائي في عصر الذكاء الاصطناعي التوليدي.
  • فهم تأثيرات الذكاء الاصطناعي التوليدي على المنظمات من حيث العمليات والهياكل. سيؤثر الذكاء الاصطناعي التوليدي بلا شك على قدرات تكنولوجيا المعلومات والأفراد، مما يؤدي إلى تغييرات في العمليات والهياكل التجارية.
  • فهم التأثيرات بين المنظمات. عندما يتخلل الذكاء الاصطناعي التوليدي المنظمات، ستختلف أنواع التفاعلات بين المنظمات، مما يؤدي إلى تغييرات في العلاقات والتفاعلات.
  • فهم المجالات التجارية الحرجة. يمكن أن تكون تأثيرات البحث أكثر أهمية في المجالات التجارية الحرجة مثل الرعاية الصحية والمالية؛ لذلك، يجب إيلاء اهتمام خاص لهذه المجالات.
  • فهم القضايا القانونية والحوكمة. يمكن أن يؤدي الذكاء الاصطناعي إلى نتائج سلبية ويؤدي إلى الجانب المظلم للذكاء الاصطناعي التوليدي، مما يتطلب دراسة القضايا القانونية والحوكمة.
  • فهم القضايا الاجتماعية الأوسع. مع تطبيق الذكاء الاصطناعي التوليدي في مختلف القطاعات التجارية، يجب دراسة قضايا اجتماعية إضافية، مثل الخصوصية والأمان.
يوفر الجدول 4 أمثلة على هذه المواضيع البحثية، التي لها أساس نظامي. ستستمر الاحتمالات الإضافية في الظهور مع استمرار تطوير هذه التكنولوجيا وتطبيقها بطرق فريدة ومثيرة للاهتمام. ستوفر العديد من فرص البحث مجالًا واسعًا من البحث لعلماء نظم المعلومات، سواء كانوا يسعون وراء البحث الفني أو السلوكي أو الإداري أو الاقتصادي.

5 الخاتمة

التقدم الأخير في الذكاء الاصطناعي التوليدي هو انتشار واختراق لأن هذه التكنولوجيا متاحة بسهولة ويمكن استخدامها مع اللغات الطبيعية دون الكثير من تدريب المستخدم. وبالتالي، فإن الذكاء الاصطناعي التوليدي لديه القدرة على التأثير على المزيد من جوانب عمليات الأعمال أكثر من معظم التقنيات السابقة. استجابةً لانفجار التطبيقات، تناولت هذه الورقة الذكاء الاصطناعي التوليدي كجيل جديد من الذكاء الاصطناعي، مما يثير قضايا تتعلق بدور التكنولوجيا (الجديدة، الناشئة، التوليدية، أو المحتملة التحول) في الأعمال والمجتمع. بالنسبة لباحثي نظم المعلومات، هناك العديد من التحديات المهمة التي تتطلب اعتبارًا دقيقًا لكل من الجوانب التقنية والاجتماعية للذكاء الاصطناعي التوليدي. اقترحت الورقة جدول أعمال لاستمرار البحث في نظم المعلومات، مع تحديد المساهمات المحتملة التي يمكن أن يقدمها هذا المجال.
الشكر الجزيل. تم دعم عمل ج. ليون زها جزئيًا من قبل المؤسسة الوطنية للعلوم الطبيعية في الصين، المنح رقم 72031001 و72231004. تم دعم عمل رومان لوكيانينكو من قبل جامعة فيرجينيا. كما تم دعم هذا البحث من قبل كلية ج. ماك روبنسون للأعمال، جامعة ولاية جورجيا. يتقدم المؤلفون بالشكر لرؤساء التحرير، الدكتور رام راميش والدكتور ه. راغاف
راو، وكذلك للمراجعين المجهولين على اعتبارهم الدقيق لهذا المخطوطة. شكرًا أيضًا لفريق إنتاج سبرينغر على جهودهم في إنتاج النسخة النهائية.
مساهمات المؤلفين. ساهم جميع المؤلفين في تصور وتصميم الدراسة. تم إعداد المواد، وجمع المراجع، وكتابة المخطوطة من قبل جميع المؤلفين بشكل تعاوني. قرأ جميع المؤلفين ووافقوا على المخطوطة النهائية.
توفر البيانات. لا يوجد مستودع بيانات يتجاوز محتوى هذه المخطوطة.

الإعلانات

المصالح المتنافسة. هذا العمل يتعلق بالبحث الأكاديمي البحت الذي لا يرتبط بأي شركة أو وكالة تجارية.
الوصول المفتوح هذه المقالة مرخصة بموجب رخصة المشاع الإبداعي النسب 4.0 الدولية، التي تسمح بالاستخدام والمشاركة والتكيف والتوزيع وإعادة الإنتاج بأي وسيلة أو صيغة، طالما أنك تعطي الائتمان المناسب للمؤلفين الأصليين والمصدر، وتوفر رابطًا لرخصة المشاع الإبداعي، وتوضح ما إذا كانت هناك تغييرات قد أُجريت. الصور أو المواد الأخرى من طرف ثالث في هذه المقالة مشمولة في رخصة المشاع الإبداعي الخاصة بالمقالة، ما لم يُشار إلى خلاف ذلك في سطر الائتمان للمواد. إذا لم تكن المادة مشمولة في رخصة المشاع الإبداعي الخاصة بالمقالة وكان استخدامك المقصود غير مسموح به بموجب اللوائح القانونية أو يتجاوز الاستخدام المسموح به، ستحتاج إلى الحصول على إذن مباشرة من صاحب حقوق الطبع والنشر. لعرض نسخة من هذه الرخصة، قم بزيارة http://creativecommons.org/licenses/by/4.0/.

References

Ackoff, R. L. (1971). Towards a system of systems concepts. Management Science, 17(11), 661-671.
Ackoff, R. L., & Emery, F. E. (2005). On purposeful systems: An interdisciplinary analysis of individual and social behavior as a system of purposeful events. Routledge.
Agrawal, S. (2023). Are LLMs the master of all trades?: Exploring domain-agnostic reasoning skills of LLMs. arXiv preprint arXiv:2303.12810.
Alavi, M., Leidner, D. E., & Mousavi, R. (2024). A knowledge management perspective of generative artificial intelligence. Journal of the Association for Information Systems, 25(1), 1-12.
Alavi, M., & Westerman, G. (2023). How generative AI will transform knowledge work. Harvard Business Review. https://hbr.org/2023/ 11/how-generative-ai-will-transform-knowledge-work. Accessed 20 Sept 2024.
Alavi, M. (2024). Generative AI and job crafting: A new frontier in knowledge work support. Harvard Business Review
Alter, S. (2015). Sociotechnical systems through a work system lens: A possible path for reconciling system conceptualizations, business realities, and humanist values in IS development. In STPIS 2015 (1st international workshop on socio-technical perspective in IS development) associated with CAISE 2015 (conference on advanced information system engineering).
Basilan, M. (2023). What are griefbots? AI-powered tech used to ‘resurrect’ the dead raises ethical questions. International Business Times. Retrieved 06/19/2023 from https://www.ibtimes.com/ what-are-griefbots-ai-powered-tech-used-resurrect-dead-raises-ethical-questions-3694800. Accessed 20 Sept 2024.
Batini, C., Lenzerini, M., & Navathe, S. B. (1986). A comparative analysis of methodologies for database schema integration. ACM Computing Surveys (CSUR), 18(4), 323-364.
Bedau, M., & Humphreys, P. (2008). Emergence: Contemporary readings in philosophy and science. MIT Press.
Bedué, P., & Fritzsche, A. (2022). Can we trust AI? An empirical investigation of trust requirements and guide to successful AI adoption. Journal of Enterprise Information Management, 35(2), 530-549.
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency. ACCT ’21 Canada March 3-10, 2021.
Berente, N., Gu, B., Recker, J., & Santhanam, R. (2021). Managing artificial intelligence. MIS Quarterly, 45(3), 1433-1450.
Bertram, E. (1972). Even permutations as a product of two conjugate cycles. Journal of Combinatorial Theory, Series A, 12(3), 368-380.
Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent Dirichlet allocation. Journal of Machine Learning Research, 3(Jan), 993-1022.
Bunge, M. (1979). Basic philosophy. Reidel Publishing Co. Dordrecht.
Bunge, M. (1996). Finding philosophy in social science. Yale University Press.
Bunge, M. (2018). Systems everywhere. In C. Negoita (Ed.), Cybernetics and applied systems (pp. 23-41). CRC Press.
Burton-Jones, A., Recker, J., Indulska, M., Green, P., & Weber, R. (2017). Assessing representation theory with a framework for pursuing success and failure. MIS Quarterly, 41(4), 1307-1334.
Cerf, V. G. (2019). AI is not an excuse! Communications of the ACM (Vol 62, Issue 10, pp. 7-7). ACM New York.
Chang, Y. B., & Gurbaxani, V. (2012). Information technology outsourcing, knowledge transfer, and firm productivity: An empirical analysis. MIS Quarterly, 36(4), 1043-1063.
Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X., Wang, C., & Wang, Y. (2024). A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3), 1-45.
Chatterjee, S., Sarker, S., Lee, M. J., Xiao, X., & Elbanna, A. (2021). A possible conceptualization of the information systems (IS) artifact: A general systems theory perspective 1. Information Systems Journal, 31(4), 550-578.
Chomsky, N. (1986). Knowledge of language: Its nature, origin, and use. Greenwood Publishing Group.
Chua, C. E. H., & Storey, V. C. (2016). Bottom-up enterprise information systems: Rethinking the roles of central IT departments. Communications of the ACM, 60(1), 66-72.
Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine Learning, 20, 273-297.
Crevier, D. (1993). AI: the tumultuous history of the search for artificial intelligence. Basic Books, Inc.
Delipetrev, B., Tsinaraki, C., & Kostić, U. (2020). Historical evolution of artificial intelligence, EUR 30221EN. Publications Office of the European Union. https://publications.jrc.ec.europa.eu/repos itory/handle/JRC120469
Demetis, D., & Lee, A. S. (2018). When humans using the IT artifact becomes IT using the human artifact. Journal of the Association for Information Systems, 19(10), 5.
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009). Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition (pp. 248-255).
Dubin, J. A., Bains, S. S., Chen, Z., Hameed, D., Nace, J., Mont, M. A., & Delanois, R. E. (2023). Using a Google web search analysis to assess the utility of ChatGPT in total joint arthroplasty. The Journal of Arthroplasty, 38(7), 1195-1202.
Dwivedi, Y. K., Kshetri, N., Hughes, L., Slade, E. L., Jeyaraj, A., Kar, A. K., Baabdullah, A. M., Koohang, A., Raghavan, V., & Ahuja, M. (2023). “So what if ChatGPT wrote it?” Multidisciplinary perspectives on opportunities, challenges and implications of
generative conversational AI for research, practice and policy. International Journal of Information Management, 71, 102642.
Feldman, M. S., & Pentland, B. T. (2003). Reconceptualizing organizational routines as a source of flexibility and change. Administrative Science Quarterly, 48(1), 94-118.
Ferrucci, D. A. (2012). Introduction to “this is Watson.”. IBM Journal of Research and Development, 56(3.4), 1: 1-1-1: 115.
French, A., Storey, V. C., & Wallace, L. (2024). A typology of disinformation intentionality and impact. Information Systems Journal, 34(4), 1324-1354.
Fügener, A., Grahl, J., Gupta, A., & Ketter, W. (2021). Will humans-in-the-loop become Borgs? Merits and pitfalls of working with AI. MIS Quarterly (MISQ), 45(3b), 1527-1556.
Fui-Hoon Nah, F., Zheng, R., Cai, J., Siau, K., & Chen, L. (2023). Generative AI and ChatGPT: Applications, challenges, and AIhuman collaboration. Journal of Information Technology Case and Application Research, 25(3), 277-304.
Gewitz, D. (2023). How to use ChatGPT to write code. https://www. zdnet.com/article/how-to-use-chatgpt-to-write-code/. Accessed 20 Sept 2024.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2020). Generative adversarial networks. Communications of the ACM, 63(11), 139-144.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014). Generative adversarial nets. Advances in Neural Information Processing Systems, 27, 1-9.
Hagendorff, T., & Fabi, S. (2022). Methodological reflections for AI alignment research using human feedback. arXiv preprint arXiv:2301.06859.
Haidt, J., & Schmidt, E. (2023). AI is about to make social media (much) more toxic. The Atlantic.
Harmon, P., & King, D. (1985). Expert systems. John Wiley & Sons. Inc.
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735-1780.
Holland, J. H. (1992). Genetic algorithms. Scientific American, 267(1), 66-73.
Jarvenpaa, S., & Klein, S. (2024). New frontiers in information systems theorizing: Human-gAI collaboration. Journal of the Association for Information Systems, 25(1), 110-121.
Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1-38.
Kajtazi, M., Holmberg, N., & Sarker, S. (2023). The changing nature of teaching future IS professionals in the era of generative AI. In In (Vol. 25, pp. 415-422). Taylor & Francis.
Kanbach, D. K., Heiduk, L., Blueher, G., Schreiter, M., & Lahmann, A. (2023). The GenAI is out of the bottle: Generative artificial intelligence from a business model innovation perspective. Review of Managerial Science, 1-32.
Kaneko, M., & Baldwin, T. (2024). A little leak will sink a great ship: survey of transparency for large language models from start to finish. arXiv preprint arXiv:2403.16139.
Ke, Y., Jin, L., Elangovan, K., Abdullah, H. R., Liu, N., Sia, A. T. H., Soh, C. R., Tung, J. Y. M., Ong, J. C. L., & Ting, D. S. W. (2024). Development and testing of retrieval augmented generation in large language models–A case study report. arXiv preprint arXiv:2402.01733.
Kelly, S. M. (2023). Snapchat’s new AI Chatbot is already raising alarms among teens and parents. Retrieved 06/19/2023 from https://edition.cnn.com/2023/04/27/tech/snapchat-my-ai-conce rns-wellness/index.html. .
Khatri, V., & Samuel, B. M. (2019). Analytics for managerial work. Communications of the ACM, 62(4), 100-100.
Kingma, D. P., & Welling, M. (2014). Auto-encoding variational Bayes. In International conference on learning representations (ICLR 2014).
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6), 84-90.
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25.
LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324.
Legner, C., Eymann, T., Hess, T., Matt, C., Böhmann, T., Drews, P., Mädche, A., Urbach, N., & Ahlemann, F. (2017). Digitalization: Opportunity and challenge for the business and information systems engineering community. Business & Information Systems Engineering, 59, 301-308.
Li, Z., Yang, Z., & Wang, M. (2023). Reinforcement learning with human feedback: Learning dynamic choices via pessimism. arXiv preprint arXiv:2305.18438.
Li, K., & Wieringa, P. A. (2000). Understanding perceived complexity in human supervisory control. Cognition, Technology & Work, 2, 75-88.
Lima, P. U., & Custodio, L. M. (2004). Artificial intelligence and systems theory: Applied to cooperative robots. International Journal of Advanced Robotic Systems, 3(1), 15.
Luhmann, N. (1995). Social systems. Stanford University Press, Palo Alto.
Lukyanenko, R., Parsons, J., Wiersma, Y. F., & Maddah, M. (2019). Expecting the unexpected: Effects of data collection design choices on the quality of crowdsourced user-generated content. MIS Quarterly, 43(2), 623-648.
Lukyanenko, R., Maass, W., & Storey, V. C. (2022a). Trust in artificial intelligence: From a foundational trust framework to emerging research opportunities. Electronic Markets, 32(4), 1993-2020.
Lukyanenko, R., Storey, V. C., & Pastor, O. (2022b). System: A core conceptual modeling construct for capturing complexity. Data & Knowledge Engineering, 141, 102062.
Luong, M.-T., Pham, H., & Manning, C. D. (2015). Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025.
Lyu, Q., Havaldar, S., Stein, A., Zhang, L., Rao, D., Wong, E., Apidianaki, M., & Callison-Burch, C. (2023). Faithful chain-ofthought reasoning. arXiv preprint arXiv:2301.13379.
McCarthy, J. (1959). Programs with common sense. In Proceedings of the Teddington conference on the mechanization of thought processes (pp. 75-91). Her Majesty’s Stationary Office.
Mei, Q., Xie, Y., Yuan, W., & Jackson, M. O. (2024). A Turing test of whether AI chatbots are behaviorally similar to humans. Proceedings of the National Academy of Sciences, 121(9), e2313925121.
Michie, D. (1963). Experiments on the mechanization of game-learning Part I. Characterization of the model and its parameters. The Computer Journal, 6(3), 232-236.
Minsky, M., & Papert, S. (1969). Perceptron: An introduction to computational geometry. MIT Press.
Motoki, F., Pinho Neto, V., & Rodrigues, V. (2024). More human than human: Measuring ChatGPT political bias. Public Choice, 198(1), 3-23.
Newell, A., Shaw, J. C., & Simon, H. A. (1962). The processes of creative thinking. In Contemporary approaches to creative thinking, 1958. University of Colorado.
O’Leary, D. E. (2022). Massive data language models and conversational AI: Emerging issues. Intelligent Systems in Accounting, Finance and Management, 29(3), 182-198.
Pachocki, J., Roditty, L., Sidford, A., Tov, R., & Williams, V. V. (2018). Approximating cycles in directed graphs: Fast algorithms for
girth and roundtrip spanners. In Proceedings of the twenty-ninth annual ACM-SIAM symposium on discrete algorithms.
Park, S. Y., Kuo, P.-Y., Barbarin, A., Kaziunas, E., Chow, A., Singh, K., Wilcox, L., & Lasecki, W. S. (2019). Identifying challenges and opportunities in human-AI collaboration in healthcare. In Conference companion publication of the 2019 on computer supported cooperative work and social computing.
Perez, C. (2010). Technological revolutions and techno-economic paradigms. Cambridge Journal of Economics, 34(1), 185-202.
. (2024). Sizing the prize. PwC’s global artificial intelligence study: exploiting the AI revolution. In What’s the real value of AI for your business and how can you capitalise?
Qiu, L., & Benbasat, I. (2010). A study of demographic embodiments of product recommendation agents in electronic commerce. International Journal of Human-Computer Studies, 68(10), 669-688.
Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language models are unsupervised multitask learners. OpenAI Blog, 1(8), 9.
Rajput, W. (2023). Integrating ChatGPT in the enterprise. https://www. linkedin.com/pulse/integrating-chatgpt-enterprise-wasim-rajput. Accessed 20 Sept 2024.
Recker, J. C., Lukyanenko, R., Jabbari Sabegh, M., Samuel, B., & Castellanos, A. (2021). From representation to mediation: A new agenda for conceptual modeling research in a digital world. MIS Quarterly, 45(1), 269-300.
Reis, L., Maier, C., Mattke, J., Creutzenberg, M., & Weitzel, T. (2020). Addressing user resistance would have prevented a healthcare AI project failure. MIS Quarterly Executive, 19(4).
Rosenblatt, F. (1961). Principles of neurodynamics : Perceptrons and the theory of brain mechanisms. Spartan Books.
Rotman, D. (2023). ChatGPT is about to revolutionize the economy. In We need to decide what that looks like. MIT Technology Review.
Russell, S., & Norvig, P. (2016). Artificial intelligence: A modern approach. Pearson.
Saba, W. S. (2023). Stochastic LLMs do not understand language: Towards symbolic, explainable and ontologically based LLMS. In International conference on conceptual modeling (pp. 3-19). ER 2023.
Sabherwal, R., & Grover, V. (2024). The societal impacts of generative artificial intelligence: A balanced perspective. Journal of the Association for Information Systems, 25(1), 13-22.
Samuel, A. L. (1960). Programming computers to play games. In Advances in computers (Vol. 1, pp. 165-192). Elsevier.
Sarker, S., Chatterjee, S., Xiao, X., & Elbanna, A. (2019). The sociotechnical axis of cohesion for the IS discipline: Its historical legacy and its continued relevance. MIS Quarterly, 43(3), 695-720.
Savage, N. (2020). The race to the top among the world’s leaders in artificial intelligence. Nature, 588(7837), S102-S102.
Savage, N. (2023). Drug discovery companies are customizing ChatGPT: Here’s how. Nature Biotechnology.
Schlindwein, S. L., & Ison, R. (2004). Human knowing and perceived complexity: Implications for systems practice. Emergence: Complexity and Organization, 6(3), 27-32.
Schmidhuber, J. (1993). Habilitation thesis: System modeling and optimization. Page 150 ff demonstrates credit assignment across the equivalent of 1,200 layers in an unfold.
Schmidhuber, J. (2015). Deep learning in neural networks: An overview. Neural Networks, 61, 85-117.
Segev, A., & Zhao, J. L. (1994). Rule management in expert database systems. Management Science, 40(6), 685-707.
Seidel, S., Berente, N., Lindberg, A., Lyytinen, K., & Nickerson, J. V. (2018). Autonomous tools and design: A triple-loop approach to humanmachine learning. Communications of the ACM, 62(1), 50-57.
Seseri, R. (2023). Generative AI: A paradigm shift in enterprise and startup opportunities. https://www.cio.com/article/474720/gener
ative-ai-a-paradigm-shift-in-enterprise-and-startup-opportunit ies.html. Accessed 20 Sept 2024.
Skyttner, L. (2001). General systems theory. World Scientific.
Smolensky, R. (1987). Algebraic methods in the theory of lower bounds for Boolean circuit complexity. In Proceedings of the nineteenth annual ACM symposium on theory of computing.
Sowa, K., Przegalinska, A., & Ciechanowski, L. (2021). Cobots in knowledge work: Human-AI collaboration in managerial professions. Journal of Business Research, 125, 135-142.
Stohr, E. A., & Zhao, J. L. (2001). Workflow automation: Overview and research issues. Information Systems Frontiers, 3, 281-296.
Storey, V. C. (2025). Knowledge management in a world of generative AI: Impact and Implications. In ACM Transactions on Management Information Systems.
Storey, V. C., Lukyanenko, R., Maass, W., & Parsons, J. (2022). Explainable AI: Opening the black box or Pandora’s Box? Communications of the ACM, 65(4), 27-29.
Susarla, A., Gopal, R., Thatcher, J. B., & Sarker, S. (2023). The Janus effect of generative AI: Charting the path for responsible conduct of scholarly activities in information systems. Information Systems Research, 34(2), 399-408.
Tang, Y., Puduppully, R., Liu, Z., & Chen, N. (2023). In-context learning of large language models for controlled dialogue summarization: A holistic benchmark and empirical analysis. In Proceedings of the 4th new frontiers in summarization workshop.
Tarafdar, M., D’arcy, J., Turel, O., & Gupta, A. (2014). The dark side of information technology. MIT Sloan Management Review.
Tesauro, G. (2002). Programming backgammon using self-teaching neural nets. Artificial Intelligence, 134(1-2), 181-199.
Thrun, S., Montemerlo, M., Dahlkamp, H., Stavens, D., Aron, A., Diebel, J., Fong, P., Gale, J., Halpenny, M., & Hoffmann, G. (2006). Stanley: The robot that won the DARPA grand challenge. Journal of Field Robotics, 23(9), 661-692.
Turing, A. M. (1950). Computing machinery and intelligence. The essential Turing: The ideas that gave birth to the computer age. Mind, LIX(236), 433-460.
Turing, A. M. (2012). Computing machinery and intelligence (1950). In The essential Turing: The ideas that gave birth to the computer age (pp. 433-464).
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.
Verganti, R., Vendraminelli, L., & Iansiti, M. (2020). Innovation and design in the age of artificial intelligence. Journal of Product Innovation Management, 37(3), 212-227.
Walters, W. P., & Murcko, M. (2020). Assessing the impact of generative AI on medicinal chemistry. Nature Biotechnology, 38(2), 143-145.
Wang, D., Weisz, J. D., Muller, M., Ram, P., Geyer, W., Dugan, C., Tausczik, Y., Samulowitz, H., & Gray, A. (2019). Human-AI collaboration in data science: Exploring data scientists’ perceptions of automated AI. In Proceedings of the ACM on human-computer interaction, 3(CSCW) (pp. 1-24).
Watkins, C. J. C. H. (1989). Learning from delayed rewards. Ph.D. Thesis, King’s College.
Weizenbaum, J. (1966). ELIZA-A computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1), 36-45.
Wu, S. P.-J., Straub, D. W., & Liang, T.-P. (2015). How information technology governance mechanisms and strategic alignment influence organizational performance. MIS Quarterly, 39(2), 497-518.
Wu, C.-J., Raghavendra, R., Gupta, U., Acun, B., Ardalani, N., Maeng, K., Chang, G., Aga, F., Huang, J., & Bai, C. (2022). Sustainable AI: Environmental implications, challenges and opportunities. Proceedings of Machine Learning and Systems, 4, 795-813.
Xue, L., Mithas, S., & Ray, G. (2021). Commitment to IT investment plans: The interplay of real earnings, management, IT decentralization, and corporate governance. MIS Quarterly, 45(1), 193-224.
Yao, B., Chen, G., Zou, R., Lu, Y., Li, J., Zhang, S., Liu, S., Hendler, J., & Wang, D. (2023). More samples or more prompt inputs? Exploring effective in-context sampling for LLM few-shot prompt engineering. arXiv preprint arXiv:2311.09782.
Yeshchenko, A., Di Ciccio, C., Mendling, J., & Polyvyanyy, A. (2019). Comprehensive process drift detection with visual analytics. In Conceptual modeling: 38th international conference, ER 2019, Salvador, Brazil, November 4-7, 2019, Proceedings (p. 38).
Yu, X., Lyu, Y., & Tsang, I. (2020). Intrinsic reward driven imitation learning via generative model. In International conference on machine learning.
Yue, W. T., Wang, Q.-H., & Hui, K. L. (2019). See no evil, hear no evil? Dissecting the impact of online hacker forums. MIS Quarterly, 43(1), 73.
Zamfirescu-Pereira, J., Wong, R. Y., Hartmann, B., & Yang, Q. (2023). Why Johnny can’t prompt: How non-AI experts try (and fail) to design LLM prompts. In Proceedings of the 2023 CHI conference on human factors in computing systems.
Zhang, B., Zhu, J., & Su, H. (2023). Toward the third generation artificial intelligence. Science China Information Sciences, 66(2), 1-19.
Zhang, J., Lan, T., Zhu, M., Liu, Z., Hoang, T., Kokane, S., Yao, W., Tan, J., Prabhakar, A., & Chen, H. (2024a). xLAM: A family of large action models to empower AI agent systems. arXiv preprint arXiv:2409.03215.
Zhang, J., Mills, D. J., & Huang, H.-W. (2024b). Enhancing travel planning and experiences with multimodal ChatGPT 4.0. In Proceedings of the 2024 international conference on innovation in artificial intelligence.
Zhao, W. (2022). Inspired, but not mimicking: A conversation between artificial intelligence and human intelligence. National Science Review, 9(6), nwac068.
Zhuang, Y., Yu, Y., Wang, K., Sun, H., & Zhang, C. (2023). Toolqa: A dataset for LLM question answering with external tools. Advances in Neural Information Processing Systems, 36, 50117-50143.
Zuboff, S. (1988). In the age of the smart machine: The future of work and power. Basic Books, Inc.
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Veda C. Storey is a Distinguished University Professor and the Tull Professor of Computer Information Systems and professor of computer science at the J. Mack Robinson College of Business, Georgia State University. Her research interests are in data management, conceptual modeling, and design science research. She is particularly interested in the assessment of the impact of new technologies on business and society from a data management perspective. Dr. Storey is a member of the AIS College of Senior Scholars and the steering committees of the International Conference of Conceptual Modeling and the Workshop on Information Technologies and Systems. She is a recipient of the Peter P. Chen Award, an ER Fellow, an AIS Fellow, and an INFORMS Fellow.
Wei Thoo Yue is a Professor of Management Information Systems in the Department of Information Systems at City University of Hong Kong. He received his Ph.D. in Management Information Systems from Purdue University. Prior to joining City University of Hong Kong, he was a faculty member at the University of Texas, Dallas. His research interests focus on the economics of information systems. His work has appeared in Management Science, Information Systems Research, MIS Quarterly, Journal of Management Information Systems, Decision Support Systems, and other journals. He currently serves as Senior Editor for Production and Operations Management.
J. Leon Zhao is a Presidential Chair Professor, Director of Center on Blockchain and Intelligent Technology, Co-head of Information Systems and Operations Management, School of Management and Economics, Chinese University of Hong Kong, Shenzhen. He was a chair professor at City University of Hong Kong and Eller Professor at University of Arizona, respectively. He has edited over 20 special issues
for academic journals including MIS Quarterly, Information Systems Research, and Journal of Operations Management. He received an IBM Faculty Award in 2005, and National Chang Jiang Scholar Chair Professorship first at Tsinghua University in 2009 and again at the Chinese University of Hong Kong, Shenzhen in 2022.
Roman Lukyanenko is an associate professor at the McIntire School of Commerce, University of Virginia. His research interests include data management and research methods (validity and artificial intelligence in literature reviews). Roman actively develops ideas, tools, and methods to improve data management and research practices. These solutions received major awards, including INFORMS Design Science Award, Governor General of Canada Gold Medal, Hebert A. Simon Design Science Award. Roman’s research appeared in Nature, MIS Quarterly, Information Systems Research, ACM Computing Surveys. His 2019 paper on quality of crowdsourced data received the Best Paper Award at MIS Quarterly.

  1. J. Leon Zhao
    leonzhao@cuhk.edu.cn
    Roman Lukyanenko
    romanl@virginia.edu
    1 Georgia State University, Atlanta, GA, USA
    City University of Hong Kong, Kowloon, Hong Kong SAR, China
    3 Chinese University of Hong Kong, Shenzhen, China
    4 University of Virginia, Charlottesville, VA, USA
  2. There could be other forms of data, but the basis is textual. Generative AI can “understand” an image because it has seen something similar before. Past images are associated with certain textual data; therefore, new images will also be associated with related textual data.
  3. The authors thank an anonymous reviewer for this positioning.
  4. Transformer-based pre-trained models approach the humanlevel benchmark with General Language Understanding Evaluation (GLUE), which is based on a collection of English language comprehension problems, rapidly. From the GLUE benchmark, a more rigorous SuperGLUE benchmark test was developed, whereby the models rapidly improved and surpassed human-level standards. From OpenAI’s own evaluations, GPT-4 performs exceptionally well on a variety of tests, including reasonings and other examinations (https://cdn. openai.com/papers/gpt-4.pdf).
  5. https://medium.com/mlearning-ai/an-ai-model-that-is-energy-effic ient-is-just-as-important-as-its-purpose-71d17822a183

Journal: Information Systems Frontiers
DOI: https://doi.org/10.1007/s10796-025-10581-7
Publication Date: 2025-02-25

Generative Artificial Intelligence: Evolving Technology, Growing Societal Impact, and Opportunities for Information Systems Research

Veda C. Storey Wei Thoo Yue J. Leon Zhao (D) ⋅ Roman Lukyanenko (D)

Accepted: 7 January 2025
© The Author(s) 2025

Abstract

The continuing, explosive developments in generative artificial intelligence (GenAI), built on large language models and related algorithms, has led to much excitement and speculation about the potential impact of this new technology. Claims include artificial intelligence (AI) being poised to revolutionize business and society and dramatically change personal life. However, it is not clear how this technology, with its significantly distinct features from past AI technologies, has transformative potential or how researchers in information systems should react to it. In this paper, we consider the evolving and emerging trends of AI in order to examine its present and predict its future impacts. Many existing papers on GenAI are either too technical for most information systems researchers or lack the depth needed to appreciate the potential impacts of GenAI. We, therefore, attempt to bridge the technical and organizational communities of GenAI from a system-oriented sociotechnical perspective. Specifically, we explore the unique features of GenAI, which are rooted in the continued change from symbolism to connectionism, and the deep systemic and inherent properties of human-AI ecosystems. We retrace the evolution of AI that proceeded the level of adoption, adaption, and use found today, in order to propose future research on various impacts of GenAI in both business and society within the context of information systems research. Our efforts are intended to contribute to the creation of a well-structured research agenda in the information systems community to support innovative strategies and operations enabled by this new wave of AI.

Keywords Artificial Intelligence (AI) • Generative AI (GenAI) • Large Language Models (LLM) • Connectionism • Societal impact • Framework for Generative AI as a Sociotechnical System • Information Systems

1 Introduction

Artificial intelligence is an immensely transformative technology, affecting organizations and individuals in ways that are impossible to fully comprehend without the benefit of a long-term historical hindsight. Artificial intelligence (AI) refers to the ability of computers to perform tasks that have historically required human cognition and other intellectual
abilities, such as perception, abstraction, inference, learning, and decision making (Russell & Norvig, 2016). AI is increasingly considered to be pinnacle of human science and engineering (Russell & Norvig, 2016). Industrial estimates are that the global economic value of AI will reach $15 trillion by 2030 (PwC, 2024). The importance of research in AI has long been recognized. It has even been suggested that the country that takes the global lead on AI may emerge as the world’s preeminent power (Savage, 2020).
Progress of AI has followed a series of significant paradigm shifts, reflecting the broader trends in technology innovation and diffusion (Perez, 2010). Before the 2000s, the field of AI underwent an “AI winter” due to stagnation in practical advances, during which AI research was primarily dominated by logic-based, model-driven learning, abstraction, and inference methods (Crevier, 1993). Since the 2000s, the ubiquity of large-scale, heterogenous data for training, the continuous expansion of computing power, and the progress in AI algorithm design, shifted the focus almost
exclusively toward machine learning (ML) and data-driven AI (Cerf, 2019), where machine learning consists of methods and algorithms that are used to make inferences from data (Goodfellow et al., 2020). With its widespread adoption, ML has begun to rapidly transform organizations and even entire industries.
Researchers in information systems (IS) have always been at the forefront of research on adopting and advancing new technologies to innovate business and social practices. Some examples of past generations of information technologies carefully investigated by IS include organizational data processing, data-driven decision-making and analytics, enterprise resource planning, and electronic commerce. We are now witnessing another AI revolution precipitated by recent developments in deep learning neural networks (Schmidhuber, 2015) and natural language processing (Rotman, 2023). Generative AI (GenAI) refers to AI systems capable of producing content such as text, images, music, programming code and other complex and creative outputs. GenAI is not new. Earlier GenAI included, for example, generative adversarial networks (Goodfellow et al., 2014). However, recent advances in GenAI are driven by the impressive performance of large language models (LLMs). LLMs are computational models that have the capability to understand and generate human language by means of a transformative ability to “predict the likelihood of word sequences or generate new text based on a given input” (Chang et al., 2024, p.2). While LLMs attempt to present human-like abilities, they also face numerous challenges due to rare or unseen words, overfitting, and complex linguistic rules. As such, new LLM architectures and training methods continue to emerge to conquer these and other challenges (Chang et al., 2024).
The result of the advances in GenAI is much interest and speculation concerning the role of artificial intelligence for everyday use, triggered, in part, by the growing popularity of tools such as ChatGPT and Dall-E (open.ai), MidJourney (midjourney.com), Google Bard (bard.google.com), and CoPilot AI (copilotai.com). ChatGPT, is an example of GenAI that can process large amounts of data to create new content (e.g., image, essay, program code, song lyrics). Fundamentally, LLMs make it possible for AI systems to assess and relate to human contexts in relatable and plausible ways (Bender et al., 2021), allowing these systems to preform knowledge-centric tasks they could not accomplish previously. Such capabilities are considered novel, with the potential to support endless new business use cases. Reviews
and analyses of GenAI tools, such as ChatGPT, range from “crossing a threshold,” where this general tool can be used for a wide range of activities (e.g., new and efficient ways to conduct work) to large-scale plagiarism and misinformation (Savage, 2023; Susarla et al., 2023).
The emergence of GenAI has led to many unanswered questions regarding how society broadly can and should respond to this new technology, particularly concerning its roles and implications. Since the debut of ChatGPT and other GenAI systems, the societal impact of GenAI has led to questions such as: How will GenAI, as technical systems transform business activities? What challenges need to be identified and resolved to ensure productive, ethical, safe, and responsible use of this technology? This gap in understanding the value and risks of GenAI leads to our main research question: What unique and valuable perspective can information systems research provide with respect to GenAI technology and its impacts on the individuals and organizations? Since research in information systems has generally been adept at studying the interconnectedness of social and technical aspects of systems (the sociotechnical nature of advances in technology), it should also be relevant for and capable of studying and contributing to the understanding of the value and risks of GenAI.
This paper introduces the concept of a “generative sociotechnical system” to explore what makes GenAI transformative and to support its ongoing diffusion in organizations and society. We develop a theoretical framework that is based upon systems theory (including a sociotechnical lens) and supported by other theoretical foundations (e.g., linguistic theory). The systemist foundation is used to position the components of GenAI and its behavior and helps to develop a research agenda for IS scholarship that builds on the existing strengths and capabilities within our discipline. The use of systems theory also enables us to study how organizations can leverage GenAI technology, how GenAI can be integrated into existing processes responsibly and profitably, and how to effectively couple this new technology with existing organizational resources. In this way, information systems scholarship should focus on exploring the nature and implications of considering GenAI as a component of broader sociotechnical systems, rather than focusing merely on the technical nature of GenAI itself.
This research makes several contributions. First, we propose a theoretical framework for understanding GenAI (Framework for Generative AI as a Sociotechnical System), from a sociotechnical perspective, consistent with existing research in information systems. Second, we show that GenAI is progressing to a point where tools will have the ability to generate unexpected results, requiring boundary conditions and new ways of considering and analyzing research initiatives. Third, we propose the notion of a generative sociotechnical system. By conceptualizing the
nature of GenAI from a systems perspective, we can develop a nuanced understanding of GenAI by leveraging existing concepts and theories of IS research. Fourth, we identify fruitful topics for future research.
This paper progresses as follows. Section 2 reviews the advances in artificial intelligence over the past three decades in order to position our understanding of the progression of GenAI. Section 3 presents a theoretical framework for understanding GenAI. Section 4 proposes potential research topics based on the framework and Sect. 5 concludes the paper.
There is a growing body of research within the information systems (IS) and business disciplines that seeks to understand and analyze the benefits and implications of generative artificial intelligence (GenAI) from various perspectives. Dwivedi et al. (2023) adopted a multidisciplinary perspective to detail the multifaceted implications of GenAI for society. Notwithstanding its promise, Fui-Hoon Nah (FuiHoon Nah et al., 2023) point to ethical, technological, and regulatory challenges associated with GenAI. Within the IS domain, research has explored the impact of GenAI on teaching (Kajtazi et al., 2023). Jarvenpaa and Klein (2024) discuss the potential of GenAI to assist with theory building
in the IS discipline. Susarla et al. (2023) suggest the potential of leveraging GenAI in various capacities to conduct scholarly work. In a broader context, Sabherwal and Grover (2024) argue that the societal impact of artificial intelligence is contingent upon its development and implementation. They emphasize the importance of considering the extent to which AI replaces versus supports humans, integrates physical and digital realities, and respects human limitations. O’Leary (2022) identifies emerging issues associated with large language models (LLMs) generally. Alavi et al. (2024) provide suggestions for how the IS discipline could focus on the role of GenAI from a knowledge management perspective, identifying related research opportunities. We continue these efforts by offering a broad and comprehensive analysis of GenAI as a technology and suggest a fruitful agenda for IS scholarship.

2.1 Prior Efforts of Research in AI and Information Systems

To assess the impact of GenAI, as well as the initiatives behind AI and its capabilities, it is crucial to understand the evolution of AI and the technological advancements that have led to its resurgence. This understanding clarifies the possibilities and limitations of AI. Figure 1 illustrates the progression of AI and the significant contributions made by researchers in information systems as AI has progressed
Fig. 1 Progression of AI and relevant research in information systems
Table 1 Summary of Early AI Breakthroughs
Year AI Artifact AI Significance Reference
1952 Checkers Demonstrated computers can learn to play at level high enough to challenge amateur human player Samuel (1960)
1955 Logic Theorist Proven 38 theorems from Principia Mathematica; introduced critical concepts in artificial intelligence (e.g., heuristics, list processing, reasoning as search) Newell et al. (1962)
1957 Perceptron Birth of connectionism; foundation of Neural Networks (NN), Deep Learning Rosenblatt (1961)
1961 MENACE First program capable of learning to play perfect game of Tic-Tac-Toe Michie (1963)
1965 ELIZA Natural language processing system imitated doctor by responding to questions similar to psychotherapist before conversation became nonsensical Weizenbaum (1966)
1969 Shakey the Robot First general-purpose mobile robot capable of reasoning, integrated research in robotics with computer vision and natural language processing Bertram (1972)
1969 The book “Perceptrons” Highlighted unrecognized limits of feed-forward, two-layered perceptron structure; fundamental shift in AI research to symbolic, disregarding connectionism Minsky and Papert (1969)
from its early work to GenAI. The part above the line shows the major stages of innovations in AI and, below the line, the relevant areas of advancements and related efforts in previous IS research. Figure 1 suggests that information systems research has always attempted to keep pace with research in AI and provides the foundation for this study on GenAI and its applications from the perspective of information systems.

2.2 Evolution from General AI to Generative AI

While GenAI is a recent development, its roots trace back to the very foundations of AI. Initial AI research aimed to create machines capable of human-like understanding and reasoning, with the famous Turing test proposed in 1950 by Alan Turing (Turing, 1950). In 1957, Herbert Simon, John Shaw, and Allen Newell developed a General Problem Solver, intending to simulate human problem solving. In 1959, John McCarthy published a paper with the telling title: “Programs with common sense” (McCarthy, 1959). The development of all-purpose systems that seek to match human cognition across a wide range of tasks later became known as Artificial General Intelligence (AGI) or General AI.
The General AI agenda proved to be less successful than anticipated, despite the high interest and investment levels (Russell and Norvig, 2016). Rather, expert systems, which were rule-based attempts to capture and represent the knowledge of human experts so it could be reused, became the predominant model for AI. The development of General AI was too difficult given the state of computing in the 80s. Rulebased and domain-specific expert systems became the choice of most AI efforts (Segev & Zhao, 1994). With LLMs, the possibility of a General AI became more imminent. For example, the company OpenAI specifically indicates that
its product, ChatGPT, can answer questions in any domain (Agrawal, 2023). This suggests that LLMs can be trained on big data that is inclusive of many domains.
To appreciate AI’s evolution from its inception to GenAI spanning 70 years of evolution from 1950s (Delipetrev et al., 2020), we recognize three primary development periods: AI Foundations (1950s-1970s); Symbolic AI (1970s-1990s); and machine learning and deep learning (1990s-2020s).
Al Foundations (1950s – 1970s). In 1950, Alan Turing published the milestone paper “Computing machinery and intelligence” (Turing, 1950) (Turing, 2012), asking the fundamental question “Can machines think?” Turing proposed an imitation game, known thereafter as the Turing test, where, if a machine could carry on a conversation indistinguishable from a conversation with a human being, it is reasonable to conclude that the machine is intelligent. The Turing test became the first experiment to attempt to measure machine intelligence. The Dartmouth conference in 1956 sparked the start of AI when McCarthy coined the term “artificial intelligence,” initiating the emergence of this new scientific field.
The primary mission of the new field of AI was to study “every aspect of any other feature of learning or intelligence [to be] be accurately described so that the machine can simulate it” (Russell & Norvig, 2016). Since then, the AI research community solved problems such as algebraic application problems, language translation, and geometric theorem proving. Table 1 summarizes important AI breakthroughs of that period, based on Delipetrev et al. (2020).
As shown in Table 1, the foundational components of large language models, namely neural networks, deep learning, and national language processing, were initiated in historical projects, Perceptron (1957). ELIZA (1961), and Shakey the Robot (1969), in the foundational period of AI. That is, it took sixty some years from the creation of AI foundations in the lab to the large language models available for general audiences.
Symbolic AI (1970s – 1990s). The era of symbolic AI focused on the development of systems based on expertcurated rules and application of logic. A major achievement of this period was the development of Expert Systems, which captured expert knowledge and represented it in a symbolic language (Harmon & King, 1985). Expert system research focused on tools for knowledge acquisition to help automate the process of designing, debugging, and maintaining rules defined by the experts. However, domain expert expertise was difficult to obtain, with expert knowledge in constant change, due to variations in regulations and environmental parameters. The period slowly descended into AI winter.
Machine Learning and Deep Learning (1990s – 2020). Modern applications of machine learning and deep learning changed the practice of research and development and began accelerating growth across various business, science, and engineering domains. The speed of AI innovation has increased, enabled by big data, expansion of computing power and new algorithms and methods. A major breakthrough was ImageNet in 2009 (Deng et al., 2009), which contained millions of annotated photos in over 20,000 categories and was critical in establishing the legitimacy of using pre-training models to train large deep neural networks. Since then, significant progress has been made in the creation of deep neural networks, which has contributed to advancements in applications such as computer vision, natural language processing (NLP) and robotics.
In 2017, Google researchers developed the transformer architecture (Vaswani et al., 2017), which later became widely adopted in the development of large language models, given its ability to process natural language. The architecture uses self-supervised learning, with a large volume of text corpora from the internet fed into the model to train it. At the heart of the architecture is the construction of selfattention from sentences (Vaswani et al., 2017) that records important information within the context of a sentence on how words relate to each other in different ways. The performance of transformer-based pre-trained models has greatly improved, with large language models distinctively enabled by changing from symbolism to connectionism.
Connectionism gained momentum in the 1990s, gradually displacing symbolism, which was based on a model of the human mind to achieve reasoning similar to that of humans (Zhang et al., 2023). Connectionism is based on connectionist modelling and systems with large networks of extremely simple processors, massively interconnected, and running in parallel, as typically found in various deep learning techniques (Smolensky, 1987). Although Rosenblatt (1961) built a perceptron in 1957-the prototype of an artificial neural network (ANN) – based on the idea of connectionism, neural networks did not achieve broad applications until the 1990s (Zhang et al., 2023).
The current AI breakthrough centers on LLMs and its associated algorithms. LLMs mimic human thinking by means of connectionism, which is rooted in machine learning via neural networks, and big data resources gathered from a variety of digital channels. Developing novel AI technologies that are safe, reliable, and extensible, requires a new, explainable and robust AI theory, thereby developing a third generation of artificial intelligence by combining the different AI paradigms (Zhang et al., 2023). Table 2 summarizes significant advances in AI leading to LLMs before 2020. Tables 1 and 2 illustrate selected milestones in the evolution of AI techniques that has led to the most recent boom of Generative AI and LLMs. It is important for IS researchers to be aware that the early AI technology (before 2017) mainly focused on specific applications such as gaming and robots. More recent AI technology (since 2017) has turned to natural language processing and generative applications. However, the concepts and algorithms of the AI technology after 2017 have evolved from the innovative efforts earlier than 2017.

2.3 AI Evolution and the Impact on IS Research

The evolution of these AI technologies has significantly impacted both sociotechnical systems and information systems (IS) research. Earlier iterations of AI artifacts were narrow and task-specific, primarily enhancing systems capabilities for specific decision-making processes. For IS researchers, this offered valuable insights into how organizations adapt to technological changes and leverage data to understand various business contexts. However, these earlier systems were limited in their scale and complexity and often lacked the ability to engage with more intricate business operations.
In contrast, the advent of GenAI has broadened its impact across diverse societal and organizational domains. GenAI is more adaptable and context-aware, thus enabling businesses to automate more complex tasks and facilitating deeper insights into decision-making processes. This transformation underscores the critical role of IS research
Table 2 Summary of Recent AI Advances
Year AI Artifact AI Significance Reference
1989 Q-learning algorithm “Learning from Delayed Rewards” improves reinforcement learning Watkins (1989)
1993 Solved “very deep learning” task Scientist solved task with over 1,000 layers in the recurrent neural network (RNN) Schmidhuber (1993)
1995 SVM success Support vector machines applied to text categorization, handwritten character recognition, and image classification Cortes and Vapnik (1995)
1997 LSTM architecture Long short-term memory (LSTM) architecture improved RNN by eliminating the long-term dependency problem Hochreiter and Schmidhuber (1997)
1998 Improved gradient-based learning Combines stochastic gradient descent algorithm with the backpropagation algorithm LeCun et al. (1998)
2002 TD-Gammon matched best player Combines neural nets and Reinforcement Learning (RL) with the self-play method Tesauro (2002); Blei et al. (2003)
2005 Stanford robot won award Drove 131 miles autonomously along an unrehearsed desert track in DARPA Grand Challenge Thrun et al. (2006)
2011 IBM Watson won Jeopardy Watson is Q&A system combining speech recognition, voice synthesis, and information retrieval, among others Ferrucci (2012)
2012 AlexNet Won ImageNet competition, possibly marking inflection point of deep learning Krizhevsky et al. (2012); Krizhevsky et al. (2017)
2014 Generative Adversarial Networks (GANs) Deep neural net architectures composed of two nets; learn to mimic distribution of data to generate content like images, music, speech, etc Goodfellow et al. (2014); Kingma and Welling (2014)
2017 Transformer DL architecture based on self-attention mechanism, important in language modelling, machine translation, and question answering Vaswani et al. (2017)
2018 OpenAI Five Defeated human team at Dota 2, complex video game messier; more realistic than Chess or Go Pachocki et al. (2018)
2019 GPT-2 Large-scale unsupervised language model; can generate coherent paragraphs of text, reading comprehension, machine translation, question answering, and summarization Radford et al. (2019)
in the adoption and use of GenAI. As highlighted in Fig. 1, IS researchers now face the challenge of exploring how these advanced systems reshape all spectrums of human activities, from organizational work processes to human-AI collaborations, while also identifying unexpected outcomes that may arise from this integration. This paper explores these developments and outlines potential research questions within IS.

3 Theoretical Framework for Understanding Generative AI

In this section we derive a theoretical framework for GenAI. The framework is intended to help understand GenAI and support proposed research.

3.1 Conceptualization of Generative AI

To understand the potential and challenges, it is important to appreciate the mechanisms for processing data inherent in GenAI. To expose these mechanisms, we examine ChatGPT, a major type and exemplar of GenAI. Note that we use ChatGPT as an instance of LLMs but the discussion is relevant to other LLMs. Figure 2 presents a conceptualization of ChatGPT, which we elaborate on next.
As shown in Fig. 2, GenAI, such as ChatGPT is an advanced question-and-answering system based on multiple computational models and big data. ChatGPT and other LLMs take a data-centric approach by training their models to generate human-like responses based on user prompts in a human language. The basic models include transformer (Vaswani et al., 2017), attention ((Luong et al., 2015), GAN
Fig. 2 Generative AI as advanced Q&A system
(Goodfellow et al., 2014, 2020), reward models (Yu et al., 2020), and RLHF (Li et al., 2023), which are used in generating responses from the texts in the big data repositories (Goodfellow et al., 2014, 2020; Li et al., 2023; Luong et al., 2015; Vaswani et al., 2017; Yu et al., 2020). It is in this sense of text transformation that LLMs are referred to as GenAI. In contrast, in conventional Q&A systems, prestored answers may be matched to keywords in the user questions using some search mechanism, thus being non-generative. Table 3 summarizes the main concepts related to large language models and their mechanisms.
We now discuss the main aspects of GenAI advancements as a unique type of information technology. This exploration is not intended to be conclusive but rather to stimulate further discussion and research from the applicational, as opposed to the purely technical, perspective.
Predictive approach – ChatGPT is a predictive approach based on deep neural networks. ChatGPT succeeds in natural language processing tasks such as chatting, translation, story making, coding, and other similar tasks in multiple languages. This is because of its representation
of human natural language in deep neural networks such as transformers. As in many machine learning techniques, ChatGPT is pre-trained based on existing texts and then used to predict outcomes based on prompt text from the user. This predictive approach means that the responses from ChatGPT are probabilistic and therefore necessarily error prone. This predictive property is easy to comprehend by technical professionals, but requires explanation for non-technical professionals.
Connectionist models – ChatGPT is generative in that it creates new connections among words via large language models. As noted, AI evolved into connectionist paradigms after exploring symbolic paradigms and then achieved great success in natural language processing via large language models, leading to ChatGPT and similar tools (Zhang et al., 2023). Neural networks essentially create simple and weighted connections among sentences and words after coding manipulations, allowing machine learning computations on hundreds of billions of inputs (e.g., text tokens). The success of ChatGPT would not have been possible because pure symbolic models are difficult to scale to big data levels. This connectionism
Table 3 Fundamental Concepts for Large Language Models
Concept Definition Year
GAN Generative adversarial networks (GAN) simultaneously train two models: a generative model that captures the data distribution and a discriminative model that estimates the probability that a sample came from the training data 2014
Attention The attention mechanism is used in machine learning and natural language processing to increase model accuracy by focusing on relevant data to enable the model to focus on certain areas of the input data, giving more weight to crucial features and disregarding unimportant ones 2015
Transformer Transformers are a type of neural network designed to handle long-range dependencies in text, capturing relationships between words and allowing a model to understand context and meaning across large sequences of inputs. For example, the relevancy and relationship between color, sky, and blue in the question: “What is the color of the sky?” lead to the output: “The sky is blue.” 2017
Reward Reward modeling is an approach in AI where a model receives a reward or score for its responses to given prompts. This reward signal serves as a reinforcement, guiding the AI model to produce desired outcomes 2020
RLHF In machine learning, reinforcement learning from human feedback (RLHF) is a technique to align an intelligent agent to human preferences 2023
can connect ChatGPT naturally with business applications via additional human and machine links to gain efficiency and accuracy. However, the connectionism in ChatGPT does not require deep knowledge of a subject or the intuition or commonsense reasoning of a human. This leaves room for improvements in reasoning and creativity. Thus, although LLMs have shown to produce coherent and human-like text, they do not ‘understand’ language in the human sense and cannot differentiate between factual from nonfactual information (Saba, 2023).
Generic capabilities – ChatGPT is generic because connectionist models do not distinguish among business sectors or knowledge domains. A single ChatGPT can serve billions of users for questions and answers, without being limited to specific contexts. This feature is a result of its large language models and gigantic neural networks of connections. However, this capability, which is based on a large-scale dataset, also suffers from low accuracy and the possibility of hallucination (that is, making up nonsensical outputs).
Complementarity with search engines – ChatGPT results complement those from search engines because the former is predictive; the latter is simply a collection of existing documents relevant to the user query. As such, ChatGPT can be considered a summary of existing documents that are more precisely related to a user query, whereas a search engine simply provides a list of search results to the user. In some way, ChatGPT offers a preliminary result to a user, who may further process the answer by ChatGPT. There is a triangular relationship amongst a user, ChatGPT, and a search engine whereby the user may determine how to use and process the results from ChatGPT and the search engine. This should be a very fruitful research topic on how to manage this knowledge processing ecosystem. The triangular relationship will not replace search engines because the results from ChatGPT might not meet all users’ needs, although some users might accept the results, without using the search engine again (Dubin et al., 2023).
Ability of understanding by ChatGPT – How ChatGPT comprehends or understands texts in comparison to humans is an important subject of research. Humans learn throughout their lifetime, starting in childhood, and enriching their brains with new knowledge without creasing throughout their entire lifetime. People express ideas, exchange opinions, and create new concepts, words, and theories. ChatGPT can process extremely large amounts of text in existing documents and generate new text, based on user requests. However, it has been noted that AI cannot understand the meaning of a context in the human sense because it does not deal with the deep semantic meaning of the words it processes, demonstrating that there is still a very large gap between artificial intelli-
gence and human intelligence (Zhao, 2022). It is simply not reasonable to assert that ChatGPT can understand or comprehend documents in human sense. There is a significant difference between the generation of new text and creation of new ideas.

3.2 Systems Theory: Traditional versus Generative AI Explained

To understand what makes GenAI transformative, we develop a theoretical framework that considers the architecture of GenAI and draws upon systems theory and other theoretical arguments (e.g., linguistic theory). These foundations help to identify the unique characteristics of GenAI and its behavior.
Many efforts to understand and design information technologies have been grounded in systems theory (Chatterjee et al., 2021). Examples include investigations into AI broadly (Skyttner, 2001), trust in AI (Lukyanenko et al., 2022a), and human-robot interactions (Lima & Custodio, 2004). Systems theory has also been previously suggested for understanding ChatGPT (Dwivedi et al., 2023). We hence adopt systems theory to understand GenAI and argue that different types of GenAI (e.g., based on large language models) are types of generative systems.
Systems theory covers related and overlapping theories that deal with the nature of systems, their interactions and uses. For general foundations, we build upon Ackoff (1971) and others, and supplement them with sociotechnical research that studies systems within social contexts (Chatterjee et al., 2021).
A system is a basic scientific and social concept: “an entity which is composed of at least two elements … each of a system’s elements is connected to every other element, directly or indirectly. No subset of elements is unrelated to any other subset” (Ackoff, 1971, p. 662). GenAI has numerous inter-connected components. Considering transformerbased LLM, there are encoder and decoder LLMs for the process of input and output systems logic. Conceptualizing GenAI as systems, requires: 1) analyzing key properties of systems, such as emergence; and 2) investigating how this technology can become a component of broader systems, especially sociotechnical systems.
Systems have two kinds of properties (Bunge, 1979, 2018): properties of parts (termed hereditary) and properties of the systems themselves (termed emergent). For example, the mass of a vehicle is a hereditary property, the sum of masses of its components. Within the context of AI, components of a neural network are the connected nodes. A hereditary property is the coefficient of a connection between nodes.
In addition, systems have emergent properties, which are properties of an entire system, rather than of any of its
components (Bedau & Humphreys, 2008). These properties emerge when the components become part of the whole and begin interacting with one another in a specific way. Since no component possesses an emergent property, the emergent properties are often not derivable from the knowledge of the properties of the components. For example, solidarity is an emergent property of a political party; no member of the party has this property. This property depends, not only on the beliefs and behaviors of the individual members (as well as extraneous factors), but also on the history, dynamics, and interactions among its members. Similarly, transparency is the emergent property of the entire neural network. While the individual neural network components may be understandable to a human, when these components are put together, they may lack transparency. Similarly, humansounding outputs of ChatGPT are emergent from the specific connectionist and natural language architecture of GenAI, which, if de-coupled and used separately will fail to render similar types of outputs (Mei et al., 2024).
Emergent properties also shape the emergent behavior of systems, analogous to the swarming behavior of a school of fish. As emergent behavior is shaped by the elusive emergent properties, emergent systems are shaped by its emergent properties, and hence its behavior. Emergence leads to qualitative, and, perhaps, ontological novelty (Bedau & Humphreys, 2008). Emergence in systems creates new realities, which do not exist at the level of their components. AI-based systems often can deliver this kind property. For instance, GAN inside GenAI, based on the concept of natural mimicry, can generate images that are deemed novel.
In general, the more complex a system, the harder it is to understand and predict its behavior. Systems theory understands complexity as being perceived and actual. Perceived complexity is a human’s interpretation and conceptualization of a system as being complex (Li & Wieringa, 2000; Schlindwein & Ison, 2004). The perception of complexity is partially and positively impacted by actual complexity. Actual complexity can be understood as the number of com-ponent-parts, along with the way in which these parts are structured and interact with one another (Lukyanenko et al., 2022b). In addition to the general systemic notions, specific considerations apply to particular types of systems (e.g., purposeful, adaptive, organic, artificial, self-organizing, selfreflective, concrete, conceptual) (Ackoff & Emery, 2005). Highlighting the distinctions allows a deeper understanding of GenAI.
Further, we note the distinction between concrete and conceptual systems. Concrete systems are systems made of material (e.g., physical) components (Bunge, 1996). Concrete systems may directly interact with other systems and change as their material components harbor energy, which can respond to and trigger change. Computers and humans are concrete systems. Organizations, which are social
systems are also concrete, since they are made of concrete components – humans and their artifacts (Luhmann, 1995).
Conceptual systems are abstract ideas bound together in the mind of a concrete system (e.g., human being) via mental rules, such as logic (Bunge, 1996). Equations, theories, hierarchies, frameworks, language grammar, logically constructed textual narratives (e.g., essays, paragraphs) are conceptual systems. Unlike concrete systems, conceptual systems do not harbor energy and are generated by concrete systems, which expend energy to create, store, modify, and communicate these systems.
Another important type of system is an adaptive system. In contrast to hard-wired or rigid systems, adaptive systems are capable of responding to environmental changes by reconfiguring their internal states. A subset of such systems are complex adaptive systems, defined as “systems composed of interacting agents described in terms of rules. The agents adapt by changing their rules as experience accumulates” (Holland, 1992, p. 10). Complex adaptive systems include natural organisms and artificial systems, such as complex machinery, including those based on AI. Such systems commonly rely on feedback loops, wherein the outputs of the system become its inputs, and hence can modulate or amplify the system’s behavior. An important consequence of adaptivity is increased difficulty in anticipating and predicting the behavior of such systems (Holland, 1992).
The objective of a GenAI process is to produce novel, complex, and self-contained outputs, in contrast to traditional machine learning, which focuses primarily on learning decision boundaries based on patterns extracted from data (Walters & Murcko, 2020). There is a profound qualitative difference between GenAI and traditional, discriminative type systems. Traditional AI technologies are fundamentally decision models, mathematical structures that seek to connect inputs to outputs in a relatively straightforward manner. Although opaque and complex, their operating principles resemble that of a measuring tape or a calculator. Often, the solution or output space is bounded, making them focus on specific tasks, therefore considered to be narrow AI.
In contrast, GenAI, which is also based on billions of iterations of training, and expressed as complex mathematical structures, determines the output in a much less straightforward way. Rather than rules that connect input to output, GenAI is a set of parameters that guides the development of a self-contained output based on an input request. A useful analogy is natural language systems.
A prevailing view in linguistics is natural languages are guided by universal grammar (UG) – the instinctive principles are instantiated when a speaker learns a particular language and uses it (Chomsky, 1986). The UG sets general principles and parameters (e.g., every word can be identified with a linguistic category). Hence, UG is a meta-language. The result of UG principles and parameters is the
generativity of human languages. For a goal, infinitely different expressions can be generated. Also, following UG, the outputs of natural language are coherent and self-contained. For example, a sentence, a paragraph, or an essay all have internally consistent and coherent structure, guided by the principles and parameters of UG.
In much the same way as natural language and other generative systems (e.g., number system), GenAI can produce potentially unlimited outputs, based on similar inputs (under the common assumption that the parameter of temperature is above zero). While these outputs are ultimately grounded in the data, novelty is in the new connections and transformation of these data. Given the extremely large set of parameters and data sources, it is nearly impossible to predict every possible way that GenAI may connect an input to an output. Although the comparison of tools, such as ChatGPT, with human creativity remain controversial, GenAI promises significantly higher level of output novelty compared with most other forms of AI.
While this may suggest it is impossible to fully control GenAI, systems theory and linguistic theory offer insights on how to potentially predict and manage these types of technologies. Specifically, the principles and parameters of GenAI are important. The linguistic structures of a language are learned by the system, mapped to another language or images, and eventually applied for different tasks; for example, a text instruction to generate a Python script.
Due to their generative ability, AI systems, such as ChatGPT, exhibit much greater complexity than traditional machine learning systems (Bender et al., 2021). These systems are also capable of relating and combining data types that go beyond text (e.g., image and speech), and, as a result, are expected to have a significantly larger number of emergent properties, leading to a variety of emergent behaviors. Following systems theory, the ability of AI systems to predict behavior (relationship of inputs to outputs) in generative systems), is considerably reduced as compared to traditional machine learning systems.
Systems theory permits other notable insight into generative systems. Not only is GenAI a system itself, but its inputs and outputs are often systems (not just system components). For example, text-to-image systems, such as DALL-E, combine LLMs and diffusion models. The prior constructs the understanding of images connected to the input texts; the latter follows a probabilistic model to synthesize images similar to the input texts. The two models were previously applied independently in systems focusing on respective tasks.
Previous generations of AI operated with only system components as their inputs and outputs. A typical machine learning model accepted inputs of a particular predefined format (controlled extraneously by the user interface), and generated outputs, which, strictly speaking, were
not usable by themselves, but required interpretation and integration into larger (conceptual or concrete) systems. For example, a machine learning model could generate a credit risk for a customer, based on a predefined feature vector corresponding to parameters of a particular customer. However, the results such as , must be interpreted and integrated into a larger conceptual system.
Generative AI, however, can accept and generate component systems, as well as systems in their own right. Indeed, when ChatGPT writes an essay, creates a Python script, or answers a question; when DALL-E or MidJourney paints an image; or when MusicLM creates music, the result is a standalone conceptual system. This system no longer needs to be embedded in another system to be usable. It could be directly used, for example, to listen to music, enjoy an image, or submit an essay as a class assignment (which, of course, poses fundamental ethical and pedagogical questions).
Finally, the ability to produce such a wide range of outputs, both component systems as well as systems in their own right, allows for a much greater diversity of the utilization of GenAI within organizations. As previous artificial intelligence, GenAI can be embedded in other technologies by harnessing its components system outputs in order to derive a coherent broader system. For example, an existing decision support system within an organization can consult GenAI for some tasks and present an output to a user which is in part based on the responses from a GenAI technology. In addition, GenAI can be integrated more directly within organizational structures, because it produces standalone conceptual systems. For example, organizational employees can use this technology in order to prepare reports, write essays, or conduct a review of a domain. Given these diverse capabilities, an important question is how to appropriately leverage this systemic versatility of GenAI.
Based on systems theory and linguistics outlined above, we propose that GenAI has three generative properties that distinguish it from the machine learning and artificial intelligence technologies of the past.
  • Strong Emergence. Whereas previous AI systems might possess emergent properties, GenAI has a particular significant ability to behave in a manner that is not directly derivable from the properties of its components and is very distant from them. This is because the outcomes of GenAI are the result of transformations from the prompts in combination with the complex knowledge of a GenAI system. Strong emergence can have both positive (e.g., ability to generate creative content) and negative consequences (e.g., difficulty to control GenAI and assure it does not harm or disadvantage people).
  • Generative Novelty. Rooted in strong emergence is generative novelty. Generative AI has the capacity to produce both expected and unexpected outputs, based on a given
    input. The outputs are products of billions of parameters tuned through billions of iterations over big training data. While these outputs are ultimately grounded in training data, the novelty lies in new ways to transform data and identify unseen patterns.
  • Systemic inputs and outputs. Generative AI has the ability to accept and produce coherent, self-contained outputs (such as self-contained responses, essays, images, animations, music). Effectively, GenAI produces self-contained conceptual systems (as opposed to snippets of decision rules), along with system components. That is to say, the systems as outputs are mapped from the system inputs based on the sophisticated transformation algorithms.
These three properties lead to new opportunities for human-computer interactions. Their organizational effects reshape the possibilities for IS research in ways that are unprecedented in the history of information systems.

3.3 Sociotechnical Perspective

Research in information systems is highly focused on sociotechnical issues where technical components are considered within the context of broader social systems (Sarker et al., 2019). Because the sociotechnical perspective fundamentally originates in the systems theory, we used to understand the nature of GenAI, it is reasonable to consider Gen AI from the sociotechnical perspective.
Figure 3 presents a Framework for GenAI as a Sociotechnical System. Because it is based on the systems perspective and follows the sociotechnical perspective, we suggest that IS research should consider the technical properties
of GenAI technology within the context of other systems, organizations, individuals, and processes. Specifically, the distinct foundational properties of GenAI of strong emergence, novelty and systemic inputs and outputs, produce new challenges and opportunities for investigating the design, use and impact of this technology in organizational and societal contexts. Such investigation permits information systems users to adopt and leverage the extensive existing knowledge of organizations and systems and to integrate this knowledge with the new fundamentals of GenAI. For example, with extensive information systems research on e-commerce and personalization, the integration of GenAI in e-commerce platforms (e.g., the Bing search engine) creates opportunities to investigate: the role of generative artificial intelligence in user adoption; trust towards e-commerce technologies; and the ability to facilitate commerce and information exchange. Because of the three fundamentally new properties of GenAI (See Fig. 3), this technology can be coupled in novel ways with the broader systems. This can lead to a research opportunity to study the intersection of the technical aspects of GenAI and the ways in which these capabilities succeed or fail to support, augment, and enhance the systems they become part of, or interact with.
A sociotechnical system based on GenAI should behave consistently with human values (AI alignment) (Hagendorff & Fabi, 2022). The alignment involves precisely prioritizing fundamental human values within the LLM. To represent these values in an AI system, humans are often involved in fine-tuning an LLM using reinforcement learning from human feedback (RLHF). Researchers in information systems should consider the broader presence of a system, including the boundary conditions that define its presence and impact within the context of an individual, organization, or society.
Fig. 3 Framework for Generative AI as a Sociotechnical System
Adopting the sociotechnical perspective, we suggest that there are broad implications of how GenAI has the potential to transform organizations and society. The first is the disruption and impact on applications that will dramatically improve human and organizational productivity. As GenAI technology advances lead to new possibilities, it will assist in the development of creative and innovative solutions. Second, a GenAI system will foster closer human and systems relationships that span both personal and business spaces. This leads to research issues focused on human-AI collaborations. Third, the versatile nature of GenAI presents numerous uncertainties, some of which may provide new opportunities but also pose threats to individuals, organizations, or society, which we consider to be the dark side of GenAI. Generative AI systems, thus, require a new design paradigm. Potential topics of designing systems with GenAI include societal concerns related to job automation and displacement of human workers, which can lead to demands for new skill sets and retraining.

4 Research Opportunities for Information Systems Research

This section proposes research themes and topics for future research on GenAI for information systems and related disciplines. They build on the sociotechnical theoretical foundations that prioritize the intersection of the technical properties of GenAI and its social entities (individuals, organizations, society).

4.1 Disruption and Impact

Potentially, GenAI could substantially enhance individual and organizational productivity, causing disruptions in numerous industries. IS researchers have a long track record of studying such impacts of new technologies.
Productivity Enhancement and Task Automation Machine learning algorithms automate data classification and prediction tasks, allowing more efficient data processing and analysis. Robotic process automation can automate routine business processes, such as data extraction and data integration. Because GenAI can process language, it can automate a wider range of jobs, leading to new use cases. For example, ChatGPT is affecting advertising because of how well it can prepare marketing and advertising messages. AI systems can free up people to perform more fulfilling or
creative work. This can lead to better efficiency, faster deci-sion-making, and higher total productivity. IS researchers have traditionally aimed to quantify the value of technology under various circumstances (Chang & Gurbaxani, 2012). Similarly, there is a need to assess the value introduced by GenAI within the greater business context.
Transformation of the Workforce and Organization The adoption of GenAI may necessitate a transformation of the workforce (Alavi, 2024; Alavi et al., 2024). Some job roles and even occupations may become obsolete or less relevant, with new roles and skills in demand. To recognize the value of AI, for instance, organizational change may be required (Reis et al., 2020). The role that senior management plays in the organization is also of interest to researchers. The capacity of GenAI to automate and optimize business processes has the potential to revolutionize businesses. Using machine learning and natural language processing can streamline operations and decision-making (Kanbach et al., 2023). Therefore, IS researchers should continue to address questions regarding organizational technology adoption.
Governance and Policy Implications. Given the disruptive potential of GenAI, the rules of engagement with new technology may change. There are many issues specific to IS, such as defining the governance model of new technology that may possess very different characteristics than those of traditional systems. This can affect an organization’s strategic orientations for managing AI-driven systems. Included are research opportunities involving use case scenarios, stakeholder engagements, and risk management approaches. Governance and policy mechanisms will invariably influence organizational performance (Wu et al., 2015). Moreover, IS scholars have contributed to understanding the interactions of these mechanisms in system’s use with the presence of internal and external organizational factors (Xue et al., 2021). Significant research is required to study the impact of systems built with the support of GenAI on organizational agility and strategy. Empowered users and IT departments supported by GenAI should be able to develop and maintain information technologies at a faster pace than in the past. This raises the question of what organizational capabilities are needed to match the expected increase in information systems adaptation and evolution. IS scholars can collaborate with business communities to study IT-strategy alignment, organizational agility, and flexible organizational routines and processes (Feldman & Pentland, 2003).

4.2 Human Machine Collaboration

Human-AI Collaboration is leading to new research issues related to fairness, unintended consequences, and regulations for human-machine collaboration. Although many
academic disciplines study human-AI collaboration, information systems have a natural capability and responsibility to help make progress in areas such as: design of humanAI systems; human and AI behaviors; economic issues of AI-assisted businesses and societies; and organizational issues of AI-intensive firms. Sample topics in human-AI collaboration can include the following:
Human and AI Agency Generative AI systems may facilitate humans to be the delegator; systems play the role of supporters or enablers. Although this role reversal has been hypothesized previously (Demetis & Lee, 2018), GenAI dramatically increases its scope and scale. We must define and explain the AI and human responsibilities in collaboration, including task responsibility, delegation, and performance evaluations. As systems mature, they may also be able to carry out the owner’s plans (or agendas). Unlike past recommendation systems, the emphasis of new AI systems might be on trying to change what people like and making a corresponding recommendation. Contexts, user engagement, recommendation systems, and systems adoption are often considered separately, but a more comprehensive consideration from IS researchers will be required.
Human-AI Interface For a user, emerging challenges, such as prompt engineering, are interface-related issues (Zamfire-scu-Pereira et al., 2023). Traditionally, interface design challenges have been concerned with making human use easier. This might be accomplished by structuring the interfaces based on predetermined choices or a flexible format (Lukyanenko et al., 2019). Making a selection from an existing list is more straightforward than typing text. Future research can investigate design approaches for GenAI systems that emphasize more open, flexible and dynamic communication process between humans and systems. However, these communication processes are dependent on how the LLM is formed, which is opaque to users.
Despite being trained with text that could reflect human values, prompt communication does not share the same communication process as humans. For example, AI systems often require more specific contextual information to accurate responses because they do not have the same reasoning ability as humans. Studies on prompt engineering could be aimed at determining the best input approaches that could be fed into the systems to interface with humans more effectively (Yao et al., 2023). Furthermore, not all humans will be comfortable with, or know how to pose the right questions to obtain the desired answers. The generative aspect of AI, could shift the conversation into topics not desired or wanted by users, followed by misconstrued prompts. Research is required on how to manage the openness and fluidity of the system-human interaction to ensure positive outcomes for users and organizations.
Human-AI Configuration Information systems researchers have rich experiences in the analysis and design of business processes in the past, leading to various models of process and workflow design (Stohr & Zhao, 2001). In the era of GenAI, process modeling will need to consider the increasing use of AI agents in place of humans, referred to as human-AI configuration (Berente et al., 2021). This will require the development of theories and principles of system design based on the understanding of human and AI behaviors in a collaborative environment where human and AI agents share information and responsibilities under emerging regulations, governing acting human agents and other humans responsible for developing and controlling the AI agents (Park et al., 2019). There will be a new area of research and development with interesting and challenging issues. It is likely that this human-AI configuration will require interdisciplinary expertise from multiple fields, such as artificial intelligence, information systems, computer science, software engineering, industrial engineering, management and economics, and law.
Design principles of human-AI systems The study of humanAI collaboration will likely lead to theories and guidelines that will be instrumental in the analysis and design of AIintensive systems in business (Fügener et al., 2021). There will be multiple aspects of design beyond human-AI configuration. As such, principles of design, development, and application of human-AI systems will be needed to facilitate the training of new talent in related disciples, development, and implementation of AI-intensive systems in business, and the creation of a new branch in the IT sector. These design principles will require new theories in human behaviors in the presence of AI agents, since the benefits and constraints of adopting AI agents will have an impact on human decisions in human-AI systems (Wang et al., 2019). The likely result will be research topics such as: mixed behavior under human and AI collaboration; reward and risk of decision making in human-AI systems; and process management in human-AI collaboration. Thus, the emergent applications in business are resulting in more attention on the research area of human-AI collaboration (Sowa et al., 2021).

4.3 Dark-Side of Generative AI

Throughout the development of information systems, researchers and practitioners have expressed concern about the unintended consequences of GenAI systems (Tarafdar et al., 2014). It is commonly believed that technology is neutral and its impact is determined by how users employ it (Yue et al., 2019). However, Generative AI is a highly adaptable technology. For people with devious purposes, there could be numerous negative consequences.
Intellectual Property Right Infringement The potential for infringement of intellectual property rights is a significant challenge. The technology has the capacity to generate novel content based on copy-protected data. For example, it has rapidly created music that imitates the style of artists, as well as voice covers that imitate their voices precisely. This concerns human artists, whose artistic talent could be “cloned” or stolen, thereby interfering with their ability to monetize their work (Yeshchenko et al., 2019). Such technology can lead to ethical and trust issues and can facilitate the creation of “digital twins” of deceased individuals, complete with their voices and communication styles (Basilan, 2023), again raising ethical concerns.
Misinformation and Deepfakes The creation of false digital information, such as images, content, and voice, that appears authentic, but is actually fake, is concerning. These deepfakes have the potential to be used for misinformation or fraudulent purposes by malevolent actors (French et al., 2024). This necessitates oversight, corrective measures, and governance (French et al., 2024). Similar to the impact of email technology on the proliferation of SPAM, the democratization of the technology has lowered the barrier to creating deepfakes, resulting in an abundance of content, making it difficult to differentiate authentic and fake information (Haidt & Schmidt, 2023).
Emotional Manipulation and Deception Programs can produce responses that appear human-like and might even be mistaken as showing personality, which can influence more advanced forms of communication, including intentions. For example, technology could be designed to keep people on websites and encourage them to buy something, by recommending products or services based on what people might like or need. This scenario has been assumed in prior IS scholarship (Qiu & Benbasat, 2010). The intuitiveness and human-like responses of GenAI, however, could give humans a deceptive sense of emotion, presence, and consciousness, possibly leading to deeper emotional connections and allowing for manipulation and deception. There could be issues associated with whether minors are emotionally ready to engage these tools (Kelly, 2023).
Hallucination and Biases Despite many impressive capabilities, GenAI suffers from several additional shortcomings and limitations. Biases and hallucinations, in particular, have been identified as an especially difficult issue to overcome. The “hallucination” problem for large language models (Ji et al., 2023) means that GenAI has a tendency to generate nonsensical outputs based on specified inputs. Related challenges are biases, which can stem from using data that does not represent reality or systematically distorts it, or improper training with the data, which may introduce additional or
compound existing distortions. Tools such as ChatGPT, have been shown to exhibit a variety of such biases, including political, moral and cultural (Motoki et al., 2024). Hence, modern GenAI appear to be the biased towards the Democrats in the US, Lula de Silva in Brazil, and the Labour Party in the UK (Motoki et al., 2024). If left unaddressed, such biases can result in erroneous decisions based on these tools and may undermine trust in particular GenAI tools, their vendors, or the entire artificial intelligence industry.
Regardless of its impressive performance, GenAI lacks the human ability to understand the meaning of its inputs (e.g., prompts of ChatGPT) and outputs (e.g., essay written in response to the prompt). Generative AI is driven by statistical probabilities of words and, more generally, pattern co-occurrences, irrespective of their actual real-world meaning (Bender et al., 2021). Large language models are merely based on probabilities of a particular word, or sentence (or more generally, token) being appropriate for particular context. Unlike humans that have imported experiences, the tools fundamentally lack ability to relate these tokens to human feelings, and thoughts and bodily experiences. The result might be potentially dangerous recommendations, because the technology lacks the ability to understand the human context around these recommendations (Storey et al., 2022). The sociotechnical lens of our paper suggests a perspective on investigating the hallucination and biases issue. On the one hand, design science researchers can contribute to the interventions that ameliorate these biases by conducting technical research. For example, data management scholars can investigate ways to evaluate biases in existing data and devise ways to procure additional data to make it more representative. Machine learning scholars can contribute by developing algorithms for handling bias to data through statistical techniques and additional processing. On the other hand, researchers should study the impact of hallucination and biases on the ways GenAI is used by people and is integrated into organizational routines. For example, an important research opportunity is understanding the boundary conditions for using GenAI for mission critical tasks. Similarly, an important research question is how to mitigate some of these biases by using additional, supporting technologies, which may be free of hallucinations and biases (such as vetted knowledge databases).
Energy Usage and Environmental Impact Another issue is that GenAI tools demand a great deal of resources (energy). Machine learning is having a staggering environmental footprint (Wu et al., 2022): the cost to train ChatGPT-3 was approximately 936 MWh , enough to power close to 100 homes for a year. Information systems scholars can
investigate the technical solutions to the energy impact problem, by devising algorithms that minimize energy usage. At the same time, a major research opportunity is understanding the full, immediate and delayed impact of widespread usage of these technologies on the environment.
Transparency As discussed, GenAI appears to be significantly less transparent than the traditional AI approaches of symbolic models. Large language models contain billions of parameters making it impossible for humans to fully understand how they make their decisions (Bender et al., 2021). Transparency of AI has long been a major barrier for trust and organizational adoption of AI technology (Bedué & Fritzsche, 2022). It is incredibly challenging to understand the logic behind LLMs, which raises significant concerns about potentially undetected biases, trust, and reliance on these systems, especially in sensitive applications (Kaneko & Baldwin, 2024). An important research opportunity is investigating the extent to which relying on such fundamentally opaque models may be unsafe, especially in particular scenarios.
The rapid adoption and incorporation of AI technology in society can result in the transformation or extinction of industries and occupations, causing workforce disruptions. It will be necessary to understand the potential repercussions to ensure that the largely opaque AI technology is aligned with human values and interests.

4.4 Designing Systems with Generative AI

Information systems were created to increase organizational efficiency and effectiveness by augmenting productivity and automating processes across various organizational functions (Zuboff, 1988). Generative AI will undoubtedly accelerate this trend. Moreover, this new technology solidifies the cloud paradigm, which promotes collaboration and integration of IT operations and services (Rajput, 2023; Seseri, 2023). Generative AI is being applied to information systems development for tasks including writing programming code (e.g., Python, JavaScript, HTML or VBA) and producing digital artifacts or components of information systems (e.g., videos, images) (Gewitz, 2023).
Information Systems Development A challenge is the rethinking of the role of users as developers. The ability of GenAI to develop systems raises fundamental questions about the process of information systems development. The redefined nature and role of end-users as developers is likely to accelerate the already existing trend where non-IT professionals significantly increase their engagement with IT development (Legner et al., 2017). A new kind of user, “empowered users,” refers to the diverse and heterogeneous
group of non-IT professionals who are motivated to take autonomous initiative and action to implement a desired change using information technology. Organizational nonIT employees increasingly implement their own solutions, such as workarounds to existing systems (Alter, 2015) and completely new solutions, especially in areas such as analytics (Khatri & Samuel, 2019). User empowerment is occurring, not only within business organizations, but also within society broadly, as people leverage GenAI’s capabilities to develop websites, apps and digital media. This bottom-up development by empowered users challenges the relationship between IT and users (Chua & Storey, 2016), leading to new avenues for research on how to best support such users.
Implementing Generative Systems The development model shifts the traditional human-centric business process to a human-in-the-loop process. There are, thus, research opportunities to redefine the role of humans and systems in the business process. Generative AI gives new impetus to the notion that a human is becoming an artifact, shaped by IT (Demetis & Lee, 2018). The systems are multi-model in nature, which is closer to how humans operate.
Many research opportunities exist in knowledge representation in the synchronization of text, image, video, and sound, based on business tasks. The transformation of different data forms to knowledge is made possible by GenAI systems, but requires new ideas of how businesses bridge proprietary data with the new technology. Another question is the boundary conditions of systems development with the support of GenAI. Some industries, such as manufacturing, construction and entertainment, have already become users of automated systems development (Seidel et al., 2018; Verganti et al., 2020). An important opportunity is understanding for which settings and industries GenAI-supported development can be effective, and where it may result in net negative outcomes. For example, design automation with GenAI may pose increased risks in mission critical and highly sensitive environments, as well as heterogeneous environments. An important design issue is how to take advantage of the capabilities of GenAI in the settings, while ensuring the safety, comfort and well-being of those affected by these designs. Greater human involvement and human-in-the loop in these contexts could be one of the solutions.
Data and Technology Integration The rapid explosion of data from multiple, heterogeneous sources presents a significant challenge. Traditionally, data integration focused on matching data from well-structured sources (e.g., database schema mapping (Batini et al., 1986)). However, the approach has limitations in ensuring the completeness and accuracy of responses. Combining disparate sources from different domains with varying levels of quality requires integrating additional techniques and technologies to
fine-tune the AI systems’ performance. Retrieval augmented generation (RAG) offers a promising approach by providing LLM with additional domain-specific knowledge to improve output performance (Ke et al., 2024). Similarly, in-context learning (ICL) provides LLMs with relevant examples and instructions through prompts. This approach instills a chain of thought during response generation, leading to a system’s ability to coherently utilize information and knowledge learned in real time (Tang et al., 2023). Researchers in information systems should focus on research questions that explore how to most effectively implement these augmenting methods in knowledge production and management in organizational contexts. This could include topics related to leveraging traditional to information systems approaches for capturing structured domain knowledge, such as through conceptual models. Knowledge representation is a fundamental and traditional part of research in information systems (Burton-Jones et al., 2017; Recker et al., 2021), making the IS community well-positioned to support RAG, ICL and other, similar initiatives.
Another research opportunity is supporting large language models with external systems that provide specific capabilities LLMs presently lack. Researchers and companies globally are rapidly responding to the opportunity to enhance the core of large language models with additional capabilities. For example, Lyu et al. (2023) propose using external solvers for math and reasoning tasks. Popular tools as well as development projects include systems that support query formulation, especially permitting drawing from personalized sources, such as corporate databases. Other tools offer reliable calculations (e.g., based on WolframAlpha API), and data visualizations (Zhuang et al., 2023). Tools emerge that integrate LLMs with external knowledge systems and platforms (e.g., travel websites permitting creation of complex itineraries based on highly personalized scenarios) (Zhang et al., 2024a, b)}. Promising research questions for information systems scholars include identification and exploration of opportunities for integrating LLMs with other, complementary IS systems. Information systems integration has been an important topic in IS, such as in the context of enterprise resource planning and social media. Another opportunity is the development of new tools that enhance LLMs, leveraging the vibrant design science IS tradition.
An emerging AI frontier is Large Action Models (LAMs), which seek to integrate the insights from LLMs with the action capabilities of autonomous agents, permitting intelligent action (Zhang et al., 2024a, b). This will allow, for example, one to directly book a vacation based on the complex itinerary created with the use of an LLM. In addition to the exploration of potential benefits of LAM, and development of LAM systems, giving direct autonomy to LLMs opens a host of questions of process, economics, trust and ethics of delegation of agency to AI.
Finally, generative AI advances the theory of information systems development by ushering in a new kind of information systems: meta-information systems. These are systems that could, in principle, take greater agency in IT development and monitoring, creation, and maintenance of other information systems. Indeed, GenAI already has capabilities that are important in information systems development; for example, the ability to analyze and structure textual documents. Hence, these tools can be used to augment and potentially automate requirements elicitation and analysis. The ability to produce database schemata and programming code can also be used to develop software components, such as database and application code and user interfaces. In this way, GenAI can become a type of meta-information technology; that is, a technology that permits the development and maintenance of other systems. Although we lack a theory of meta-information systems, GenAI could become an important use case.

4.5 Significant IS Research Topics

Technology has evolved significantly over time. In the history of technological transformation, agricultural mechanization liberated over of farmers in developed countries. Similarly, robotic automation could release the majority of blue-collar workers from factory flowlines. Perhaps GenAI and associated business applications could release whitecollar workers from office desks for many conventional processes and tasks, thereby changing the structure of their jobs and how they acquire knowledge (Alavi, 2024; Alavi & Westerman, 2023; Storey, 2025). If so, how can information systems researchers support this change? The research challenges and topics discussed in this paper are exemplary and the exploration of research frameworks, theories, and directions towards achieving this goal. Clearly, information systems researchers will strive to offer contributions to GenAI as it continues to evolve.
Researchers in information systems should help shape the impact of the progression of GenAI, focusing on its sociotechnical aspects and properties, as well as the manner in which it interacts with individuals, organizations, and society. This leads to various research challenges and opportunities.
  • Understand and improve business technology based on GenAI. Design science researchers in information systems will have an opportunity to study technological issues in GenAI applications in business.
  • Understand the impacts of GenAI on individuals including workers and general users. GenAI applications and plug-ins will be increasingly used by business work-
Table 4 Research Opportunities on Generative AI for Information Systems
Challenges Research topics and examples
Understand and improve business technology based on GenAI
– Is generative AI transformative or disruptive to business? What are its application conditions during adoption?
– What representation(s) is needed (e.g., natural language, program code, algorithmic formulations) for the management of AI knowledge?
– What multimedia generation is appropriate? How should text, image, video, sound be synchronized based on tasks in business processes?
– Are there different types of GenAI applications for various users, e.g., natural language explanation (customers), algorithm specification (data scientists), and biases or unintended consequences (managers)?
Understand the impacts of GenAI on individuals including workers and general users
– How to evaluate and increase trust in GenAI and the systems with which it interacts?
– How can GenAI be used to augment existing knowledge bases?
– What are the potential negative impacts or the dark side of GenAI on human behavior, and what are the remedies?
– What are the implications for knowledge production and the amount of knowledge workers required to learn and retain?
– What is the importance of and emphasis needed for intelligent search?
– How can goodness and fairness of knowledge outcomes be assessed?
– Is it possible to identify and stop undesirable people from making connections via GenAI? If so, how?
Understand the impacts of GenAI on organizations in terms of processes and structures
– How is GenAI infused and adopted in organizations?
– What is the impact on business strategy?
– How do we manage and utilize the technology in organizations?
– What changes to managerial practices are needed?
– How do researchers model business tasks and processes that involve both human and machine agents?
– How can human and machine collaboration be optimized?
– What kind of human-in-the-loop or AI-human collaboration is required?
– What kind of interaction and supporting interfaces are needed?
Understand inter-organizational impacts of GenAI
– Can GenAI change the competitive situation and outcome of an industry; e.g., via more intelligent and dynamic pricing that might affect the market?
– How can GenAI improve productivity and efficiency; e.g., impact on organization’s knowledge base, communications, R&D, product design?
– What are the potential adverse organizational impacts of GenAI, and what are the mitigation strategies?
Understand mission-critical business domains for GenAI adoption
– What is the impact of GenAI on mission-critical domains, such as medicine, military, transportation, food industry?
– What are implications for teaching information systems and other disciplines?
– How do we integrate GenAI into educational processes?
Understand legal and governance issues of GenAI
– What governance needs to be put into place?
– What if the collective influence of a company (companies) provides the majority of the input for connectionism features?
– How should internal and external risks be managed?
– How do we choose tasks while minimizing personal risk in exposure to penalty and lawsuits?
Understand broader societal issues of GenAI
– What are new ethical challenges of GenAI?
– What is the need for reference identification and verification of plagiarism?
– What is the extent of automation and job replacement? How do we measure it? Does it create “societal stress?”
– To what extent can creativity be achieved with GenAI, particularly for teams?
– How can we prevent personal information from being stolen in the era of GenAI? How can we protect individual’s intellectual property while allowing data to be used in GenAI?
– Will GenAI change the labor markets and in what way?
ers and general users, which will require a renewal of many research topics in end user computing in the era of GenAI.
  • Understand the impacts of GenAI on organizations in terms of processes and structures. GenAI will undoubtedly affect the capabilities of IT and individuals, leading to changes in business processes and structures.
  • Understand inter-organizational impacts. When GenAI permeates organizations, the types of interactions between and among organizations will vary, leading to relationship and interactive changes.
  • Understand mission-critical business domains. Research impacts can be more significant in mission-critical business domains such as healthcare and finance; therefore, special attention should be paid to these areas.
  • Understand legal and governance issues. AI can lead to negative outcomes and leading to a dark-side of GenAI, requiring the studies of legal and governance issues.
  • Understand broader societal issues. As GenAI is applied in various business sectors, additional societal issues, such as privacy and security, should be studied.
Table 4 provides examples of these research topics, which have a systems foundation. Additional possibilities will continue to emerge as this technology continues to be developed and applied in unique and interesting ways. The many research opportunities will offer a vast area of research for information systems scholars, whether they pursue technical, behavioral, managerial or economic research.

5 Conclusion

Recent advancement of GenAI is diffusive and penetrating because this technology is readily available and can be used with natural languages without a great deal of user training. As such, GenAI has the potential to affect more aspects of business operations than most previous technologies. In response to the explosion of applications, this paper has examined GenAI as the next generation of AI, which raises issues related to the role of technology (new, emerging, generative, or potentially transformative) in business and society. For information systems researchers, there are many important challenges that require careful consideration of both technical and societal aspects of GenAI. The paper has proposed an agenda for continued information systems research, identifying the potential contributions the field can make.
Acknowledgements J. Leon Zhao’s work is partially supported by the National Natural Science Foundation of China grants No. 72031001 and 72231004. Roman Lukyanenko’s work is supported by the University of Virginia. This research was also supported by the J. Mack Robinson College of Business, Georgia State University. The authors wish to thank the Editors-in-Chief, Dr. Ram Ramesh and Dr. H. Raghav
Rao, as well as the anonymous reviewers for their careful consideration of this manuscript. Thanks also to Springer Production team for their efforts in producing the final version.
Authors’ Contributions All authors contributed to the study conception and design. Material preparation, reference collection and manuscript writing were performed by all authors collaboratively. All authors read and approved the final manuscript.
Data Availability There is no data repository beyond the content of this manuscript.

Declarations

Competing Interests This work is about pure academic research that is not related to any commercial company or agency.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.

References

Ackoff, R. L. (1971). Towards a system of systems concepts. Management Science, 17(11), 661-671.
Ackoff, R. L., & Emery, F. E. (2005). On purposeful systems: An interdisciplinary analysis of individual and social behavior as a system of purposeful events. Routledge.
Agrawal, S. (2023). Are LLMs the master of all trades?: Exploring domain-agnostic reasoning skills of LLMs. arXiv preprint arXiv:2303.12810.
Alavi, M., Leidner, D. E., & Mousavi, R. (2024). A knowledge management perspective of generative artificial intelligence. Journal of the Association for Information Systems, 25(1), 1-12.
Alavi, M., & Westerman, G. (2023). How generative AI will transform knowledge work. Harvard Business Review. https://hbr.org/2023/ 11/how-generative-ai-will-transform-knowledge-work. Accessed 20 Sept 2024.
Alavi, M. (2024). Generative AI and job crafting: A new frontier in knowledge work support. Harvard Business Review
Alter, S. (2015). Sociotechnical systems through a work system lens: A possible path for reconciling system conceptualizations, business realities, and humanist values in IS development. In STPIS 2015 (1st international workshop on socio-technical perspective in IS development) associated with CAISE 2015 (conference on advanced information system engineering).
Basilan, M. (2023). What are griefbots? AI-powered tech used to ‘resurrect’ the dead raises ethical questions. International Business Times. Retrieved 06/19/2023 from https://www.ibtimes.com/ what-are-griefbots-ai-powered-tech-used-resurrect-dead-raises-ethical-questions-3694800. Accessed 20 Sept 2024.
Batini, C., Lenzerini, M., & Navathe, S. B. (1986). A comparative analysis of methodologies for database schema integration. ACM Computing Surveys (CSUR), 18(4), 323-364.
Bedau, M., & Humphreys, P. (2008). Emergence: Contemporary readings in philosophy and science. MIT Press.
Bedué, P., & Fritzsche, A. (2022). Can we trust AI? An empirical investigation of trust requirements and guide to successful AI adoption. Journal of Enterprise Information Management, 35(2), 530-549.
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency. ACCT ’21 Canada March 3-10, 2021.
Berente, N., Gu, B., Recker, J., & Santhanam, R. (2021). Managing artificial intelligence. MIS Quarterly, 45(3), 1433-1450.
Bertram, E. (1972). Even permutations as a product of two conjugate cycles. Journal of Combinatorial Theory, Series A, 12(3), 368-380.
Blei, D. M., Ng, A. Y., & Jordan, M. I. (2003). Latent Dirichlet allocation. Journal of Machine Learning Research, 3(Jan), 993-1022.
Bunge, M. (1979). Basic philosophy. Reidel Publishing Co. Dordrecht.
Bunge, M. (1996). Finding philosophy in social science. Yale University Press.
Bunge, M. (2018). Systems everywhere. In C. Negoita (Ed.), Cybernetics and applied systems (pp. 23-41). CRC Press.
Burton-Jones, A., Recker, J., Indulska, M., Green, P., & Weber, R. (2017). Assessing representation theory with a framework for pursuing success and failure. MIS Quarterly, 41(4), 1307-1334.
Cerf, V. G. (2019). AI is not an excuse! Communications of the ACM (Vol 62, Issue 10, pp. 7-7). ACM New York.
Chang, Y. B., & Gurbaxani, V. (2012). Information technology outsourcing, knowledge transfer, and firm productivity: An empirical analysis. MIS Quarterly, 36(4), 1043-1063.
Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., Chen, H., Yi, X., Wang, C., & Wang, Y. (2024). A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology, 15(3), 1-45.
Chatterjee, S., Sarker, S., Lee, M. J., Xiao, X., & Elbanna, A. (2021). A possible conceptualization of the information systems (IS) artifact: A general systems theory perspective 1. Information Systems Journal, 31(4), 550-578.
Chomsky, N. (1986). Knowledge of language: Its nature, origin, and use. Greenwood Publishing Group.
Chua, C. E. H., & Storey, V. C. (2016). Bottom-up enterprise information systems: Rethinking the roles of central IT departments. Communications of the ACM, 60(1), 66-72.
Cortes, C., & Vapnik, V. (1995). Support-vector networks. Machine Learning, 20, 273-297.
Crevier, D. (1993). AI: the tumultuous history of the search for artificial intelligence. Basic Books, Inc.
Delipetrev, B., Tsinaraki, C., & Kostić, U. (2020). Historical evolution of artificial intelligence, EUR 30221EN. Publications Office of the European Union. https://publications.jrc.ec.europa.eu/repos itory/handle/JRC120469
Demetis, D., & Lee, A. S. (2018). When humans using the IT artifact becomes IT using the human artifact. Journal of the Association for Information Systems, 19(10), 5.
Deng, J., Dong, W., Socher, R., Li, L.-J., Li, K., & Fei-Fei, L. (2009). Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition (pp. 248-255).
Dubin, J. A., Bains, S. S., Chen, Z., Hameed, D., Nace, J., Mont, M. A., & Delanois, R. E. (2023). Using a Google web search analysis to assess the utility of ChatGPT in total joint arthroplasty. The Journal of Arthroplasty, 38(7), 1195-1202.
Dwivedi, Y. K., Kshetri, N., Hughes, L., Slade, E. L., Jeyaraj, A., Kar, A. K., Baabdullah, A. M., Koohang, A., Raghavan, V., & Ahuja, M. (2023). “So what if ChatGPT wrote it?” Multidisciplinary perspectives on opportunities, challenges and implications of
generative conversational AI for research, practice and policy. International Journal of Information Management, 71, 102642.
Feldman, M. S., & Pentland, B. T. (2003). Reconceptualizing organizational routines as a source of flexibility and change. Administrative Science Quarterly, 48(1), 94-118.
Ferrucci, D. A. (2012). Introduction to “this is Watson.”. IBM Journal of Research and Development, 56(3.4), 1: 1-1-1: 115.
French, A., Storey, V. C., & Wallace, L. (2024). A typology of disinformation intentionality and impact. Information Systems Journal, 34(4), 1324-1354.
Fügener, A., Grahl, J., Gupta, A., & Ketter, W. (2021). Will humans-in-the-loop become Borgs? Merits and pitfalls of working with AI. MIS Quarterly (MISQ), 45(3b), 1527-1556.
Fui-Hoon Nah, F., Zheng, R., Cai, J., Siau, K., & Chen, L. (2023). Generative AI and ChatGPT: Applications, challenges, and AIhuman collaboration. Journal of Information Technology Case and Application Research, 25(3), 277-304.
Gewitz, D. (2023). How to use ChatGPT to write code. https://www. zdnet.com/article/how-to-use-chatgpt-to-write-code/. Accessed 20 Sept 2024.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2020). Generative adversarial networks. Communications of the ACM, 63(11), 139-144.
Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A., & Bengio, Y. (2014). Generative adversarial nets. Advances in Neural Information Processing Systems, 27, 1-9.
Hagendorff, T., & Fabi, S. (2022). Methodological reflections for AI alignment research using human feedback. arXiv preprint arXiv:2301.06859.
Haidt, J., & Schmidt, E. (2023). AI is about to make social media (much) more toxic. The Atlantic.
Harmon, P., & King, D. (1985). Expert systems. John Wiley & Sons. Inc.
Hochreiter, S., & Schmidhuber, J. (1997). Long short-term memory. Neural Computation, 9(8), 1735-1780.
Holland, J. H. (1992). Genetic algorithms. Scientific American, 267(1), 66-73.
Jarvenpaa, S., & Klein, S. (2024). New frontiers in information systems theorizing: Human-gAI collaboration. Journal of the Association for Information Systems, 25(1), 110-121.
Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., Ishii, E., Bang, Y. J., Madotto, A., & Fung, P. (2023). Survey of hallucination in natural language generation. ACM Computing Surveys, 55(12), 1-38.
Kajtazi, M., Holmberg, N., & Sarker, S. (2023). The changing nature of teaching future IS professionals in the era of generative AI. In In (Vol. 25, pp. 415-422). Taylor & Francis.
Kanbach, D. K., Heiduk, L., Blueher, G., Schreiter, M., & Lahmann, A. (2023). The GenAI is out of the bottle: Generative artificial intelligence from a business model innovation perspective. Review of Managerial Science, 1-32.
Kaneko, M., & Baldwin, T. (2024). A little leak will sink a great ship: survey of transparency for large language models from start to finish. arXiv preprint arXiv:2403.16139.
Ke, Y., Jin, L., Elangovan, K., Abdullah, H. R., Liu, N., Sia, A. T. H., Soh, C. R., Tung, J. Y. M., Ong, J. C. L., & Ting, D. S. W. (2024). Development and testing of retrieval augmented generation in large language models–A case study report. arXiv preprint arXiv:2402.01733.
Kelly, S. M. (2023). Snapchat’s new AI Chatbot is already raising alarms among teens and parents. Retrieved 06/19/2023 from https://edition.cnn.com/2023/04/27/tech/snapchat-my-ai-conce rns-wellness/index.html. .
Khatri, V., & Samuel, B. M. (2019). Analytics for managerial work. Communications of the ACM, 62(4), 100-100.
Kingma, D. P., & Welling, M. (2014). Auto-encoding variational Bayes. In International conference on learning representations (ICLR 2014).
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2017). Imagenet classification with deep convolutional neural networks. Communications of the ACM, 60(6), 84-90.
Krizhevsky, A., Sutskever, I., & Hinton, G. E. (2012). Imagenet classification with deep convolutional neural networks. Advances in neural information processing systems, 25.
LeCun, Y., Bottou, L., Bengio, Y., & Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324.
Legner, C., Eymann, T., Hess, T., Matt, C., Böhmann, T., Drews, P., Mädche, A., Urbach, N., & Ahlemann, F. (2017). Digitalization: Opportunity and challenge for the business and information systems engineering community. Business & Information Systems Engineering, 59, 301-308.
Li, Z., Yang, Z., & Wang, M. (2023). Reinforcement learning with human feedback: Learning dynamic choices via pessimism. arXiv preprint arXiv:2305.18438.
Li, K., & Wieringa, P. A. (2000). Understanding perceived complexity in human supervisory control. Cognition, Technology & Work, 2, 75-88.
Lima, P. U., & Custodio, L. M. (2004). Artificial intelligence and systems theory: Applied to cooperative robots. International Journal of Advanced Robotic Systems, 3(1), 15.
Luhmann, N. (1995). Social systems. Stanford University Press, Palo Alto.
Lukyanenko, R., Parsons, J., Wiersma, Y. F., & Maddah, M. (2019). Expecting the unexpected: Effects of data collection design choices on the quality of crowdsourced user-generated content. MIS Quarterly, 43(2), 623-648.
Lukyanenko, R., Maass, W., & Storey, V. C. (2022a). Trust in artificial intelligence: From a foundational trust framework to emerging research opportunities. Electronic Markets, 32(4), 1993-2020.
Lukyanenko, R., Storey, V. C., & Pastor, O. (2022b). System: A core conceptual modeling construct for capturing complexity. Data & Knowledge Engineering, 141, 102062.
Luong, M.-T., Pham, H., & Manning, C. D. (2015). Effective approaches to attention-based neural machine translation. arXiv preprint arXiv:1508.04025.
Lyu, Q., Havaldar, S., Stein, A., Zhang, L., Rao, D., Wong, E., Apidianaki, M., & Callison-Burch, C. (2023). Faithful chain-ofthought reasoning. arXiv preprint arXiv:2301.13379.
McCarthy, J. (1959). Programs with common sense. In Proceedings of the Teddington conference on the mechanization of thought processes (pp. 75-91). Her Majesty’s Stationary Office.
Mei, Q., Xie, Y., Yuan, W., & Jackson, M. O. (2024). A Turing test of whether AI chatbots are behaviorally similar to humans. Proceedings of the National Academy of Sciences, 121(9), e2313925121.
Michie, D. (1963). Experiments on the mechanization of game-learning Part I. Characterization of the model and its parameters. The Computer Journal, 6(3), 232-236.
Minsky, M., & Papert, S. (1969). Perceptron: An introduction to computational geometry. MIT Press.
Motoki, F., Pinho Neto, V., & Rodrigues, V. (2024). More human than human: Measuring ChatGPT political bias. Public Choice, 198(1), 3-23.
Newell, A., Shaw, J. C., & Simon, H. A. (1962). The processes of creative thinking. In Contemporary approaches to creative thinking, 1958. University of Colorado.
O’Leary, D. E. (2022). Massive data language models and conversational AI: Emerging issues. Intelligent Systems in Accounting, Finance and Management, 29(3), 182-198.
Pachocki, J., Roditty, L., Sidford, A., Tov, R., & Williams, V. V. (2018). Approximating cycles in directed graphs: Fast algorithms for
girth and roundtrip spanners. In Proceedings of the twenty-ninth annual ACM-SIAM symposium on discrete algorithms.
Park, S. Y., Kuo, P.-Y., Barbarin, A., Kaziunas, E., Chow, A., Singh, K., Wilcox, L., & Lasecki, W. S. (2019). Identifying challenges and opportunities in human-AI collaboration in healthcare. In Conference companion publication of the 2019 on computer supported cooperative work and social computing.
Perez, C. (2010). Technological revolutions and techno-economic paradigms. Cambridge Journal of Economics, 34(1), 185-202.
. (2024). Sizing the prize. PwC’s global artificial intelligence study: exploiting the AI revolution. In What’s the real value of AI for your business and how can you capitalise?
Qiu, L., & Benbasat, I. (2010). A study of demographic embodiments of product recommendation agents in electronic commerce. International Journal of Human-Computer Studies, 68(10), 669-688.
Radford, A., Wu, J., Child, R., Luan, D., Amodei, D., & Sutskever, I. (2019). Language models are unsupervised multitask learners. OpenAI Blog, 1(8), 9.
Rajput, W. (2023). Integrating ChatGPT in the enterprise. https://www. linkedin.com/pulse/integrating-chatgpt-enterprise-wasim-rajput. Accessed 20 Sept 2024.
Recker, J. C., Lukyanenko, R., Jabbari Sabegh, M., Samuel, B., & Castellanos, A. (2021). From representation to mediation: A new agenda for conceptual modeling research in a digital world. MIS Quarterly, 45(1), 269-300.
Reis, L., Maier, C., Mattke, J., Creutzenberg, M., & Weitzel, T. (2020). Addressing user resistance would have prevented a healthcare AI project failure. MIS Quarterly Executive, 19(4).
Rosenblatt, F. (1961). Principles of neurodynamics : Perceptrons and the theory of brain mechanisms. Spartan Books.
Rotman, D. (2023). ChatGPT is about to revolutionize the economy. In We need to decide what that looks like. MIT Technology Review.
Russell, S., & Norvig, P. (2016). Artificial intelligence: A modern approach. Pearson.
Saba, W. S. (2023). Stochastic LLMs do not understand language: Towards symbolic, explainable and ontologically based LLMS. In International conference on conceptual modeling (pp. 3-19). ER 2023.
Sabherwal, R., & Grover, V. (2024). The societal impacts of generative artificial intelligence: A balanced perspective. Journal of the Association for Information Systems, 25(1), 13-22.
Samuel, A. L. (1960). Programming computers to play games. In Advances in computers (Vol. 1, pp. 165-192). Elsevier.
Sarker, S., Chatterjee, S., Xiao, X., & Elbanna, A. (2019). The sociotechnical axis of cohesion for the IS discipline: Its historical legacy and its continued relevance. MIS Quarterly, 43(3), 695-720.
Savage, N. (2020). The race to the top among the world’s leaders in artificial intelligence. Nature, 588(7837), S102-S102.
Savage, N. (2023). Drug discovery companies are customizing ChatGPT: Here’s how. Nature Biotechnology.
Schlindwein, S. L., & Ison, R. (2004). Human knowing and perceived complexity: Implications for systems practice. Emergence: Complexity and Organization, 6(3), 27-32.
Schmidhuber, J. (1993). Habilitation thesis: System modeling and optimization. Page 150 ff demonstrates credit assignment across the equivalent of 1,200 layers in an unfold.
Schmidhuber, J. (2015). Deep learning in neural networks: An overview. Neural Networks, 61, 85-117.
Segev, A., & Zhao, J. L. (1994). Rule management in expert database systems. Management Science, 40(6), 685-707.
Seidel, S., Berente, N., Lindberg, A., Lyytinen, K., & Nickerson, J. V. (2018). Autonomous tools and design: A triple-loop approach to humanmachine learning. Communications of the ACM, 62(1), 50-57.
Seseri, R. (2023). Generative AI: A paradigm shift in enterprise and startup opportunities. https://www.cio.com/article/474720/gener
ative-ai-a-paradigm-shift-in-enterprise-and-startup-opportunit ies.html. Accessed 20 Sept 2024.
Skyttner, L. (2001). General systems theory. World Scientific.
Smolensky, R. (1987). Algebraic methods in the theory of lower bounds for Boolean circuit complexity. In Proceedings of the nineteenth annual ACM symposium on theory of computing.
Sowa, K., Przegalinska, A., & Ciechanowski, L. (2021). Cobots in knowledge work: Human-AI collaboration in managerial professions. Journal of Business Research, 125, 135-142.
Stohr, E. A., & Zhao, J. L. (2001). Workflow automation: Overview and research issues. Information Systems Frontiers, 3, 281-296.
Storey, V. C. (2025). Knowledge management in a world of generative AI: Impact and Implications. In ACM Transactions on Management Information Systems.
Storey, V. C., Lukyanenko, R., Maass, W., & Parsons, J. (2022). Explainable AI: Opening the black box or Pandora’s Box? Communications of the ACM, 65(4), 27-29.
Susarla, A., Gopal, R., Thatcher, J. B., & Sarker, S. (2023). The Janus effect of generative AI: Charting the path for responsible conduct of scholarly activities in information systems. Information Systems Research, 34(2), 399-408.
Tang, Y., Puduppully, R., Liu, Z., & Chen, N. (2023). In-context learning of large language models for controlled dialogue summarization: A holistic benchmark and empirical analysis. In Proceedings of the 4th new frontiers in summarization workshop.
Tarafdar, M., D’arcy, J., Turel, O., & Gupta, A. (2014). The dark side of information technology. MIT Sloan Management Review.
Tesauro, G. (2002). Programming backgammon using self-teaching neural nets. Artificial Intelligence, 134(1-2), 181-199.
Thrun, S., Montemerlo, M., Dahlkamp, H., Stavens, D., Aron, A., Diebel, J., Fong, P., Gale, J., Halpenny, M., & Hoffmann, G. (2006). Stanley: The robot that won the DARPA grand challenge. Journal of Field Robotics, 23(9), 661-692.
Turing, A. M. (1950). Computing machinery and intelligence. The essential Turing: The ideas that gave birth to the computer age. Mind, LIX(236), 433-460.
Turing, A. M. (2012). Computing machinery and intelligence (1950). In The essential Turing: The ideas that gave birth to the computer age (pp. 433-464).
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, Ł., & Polosukhin, I. (2017). Attention is all you need. Advances in Neural Information Processing Systems, 30.
Verganti, R., Vendraminelli, L., & Iansiti, M. (2020). Innovation and design in the age of artificial intelligence. Journal of Product Innovation Management, 37(3), 212-227.
Walters, W. P., & Murcko, M. (2020). Assessing the impact of generative AI on medicinal chemistry. Nature Biotechnology, 38(2), 143-145.
Wang, D., Weisz, J. D., Muller, M., Ram, P., Geyer, W., Dugan, C., Tausczik, Y., Samulowitz, H., & Gray, A. (2019). Human-AI collaboration in data science: Exploring data scientists’ perceptions of automated AI. In Proceedings of the ACM on human-computer interaction, 3(CSCW) (pp. 1-24).
Watkins, C. J. C. H. (1989). Learning from delayed rewards. Ph.D. Thesis, King’s College.
Weizenbaum, J. (1966). ELIZA-A computer program for the study of natural language communication between man and machine. Communications of the ACM, 9(1), 36-45.
Wu, S. P.-J., Straub, D. W., & Liang, T.-P. (2015). How information technology governance mechanisms and strategic alignment influence organizational performance. MIS Quarterly, 39(2), 497-518.
Wu, C.-J., Raghavendra, R., Gupta, U., Acun, B., Ardalani, N., Maeng, K., Chang, G., Aga, F., Huang, J., & Bai, C. (2022). Sustainable AI: Environmental implications, challenges and opportunities. Proceedings of Machine Learning and Systems, 4, 795-813.
Xue, L., Mithas, S., & Ray, G. (2021). Commitment to IT investment plans: The interplay of real earnings, management, IT decentralization, and corporate governance. MIS Quarterly, 45(1), 193-224.
Yao, B., Chen, G., Zou, R., Lu, Y., Li, J., Zhang, S., Liu, S., Hendler, J., & Wang, D. (2023). More samples or more prompt inputs? Exploring effective in-context sampling for LLM few-shot prompt engineering. arXiv preprint arXiv:2311.09782.
Yeshchenko, A., Di Ciccio, C., Mendling, J., & Polyvyanyy, A. (2019). Comprehensive process drift detection with visual analytics. In Conceptual modeling: 38th international conference, ER 2019, Salvador, Brazil, November 4-7, 2019, Proceedings (p. 38).
Yu, X., Lyu, Y., & Tsang, I. (2020). Intrinsic reward driven imitation learning via generative model. In International conference on machine learning.
Yue, W. T., Wang, Q.-H., & Hui, K. L. (2019). See no evil, hear no evil? Dissecting the impact of online hacker forums. MIS Quarterly, 43(1), 73.
Zamfirescu-Pereira, J., Wong, R. Y., Hartmann, B., & Yang, Q. (2023). Why Johnny can’t prompt: How non-AI experts try (and fail) to design LLM prompts. In Proceedings of the 2023 CHI conference on human factors in computing systems.
Zhang, B., Zhu, J., & Su, H. (2023). Toward the third generation artificial intelligence. Science China Information Sciences, 66(2), 1-19.
Zhang, J., Lan, T., Zhu, M., Liu, Z., Hoang, T., Kokane, S., Yao, W., Tan, J., Prabhakar, A., & Chen, H. (2024a). xLAM: A family of large action models to empower AI agent systems. arXiv preprint arXiv:2409.03215.
Zhang, J., Mills, D. J., & Huang, H.-W. (2024b). Enhancing travel planning and experiences with multimodal ChatGPT 4.0. In Proceedings of the 2024 international conference on innovation in artificial intelligence.
Zhao, W. (2022). Inspired, but not mimicking: A conversation between artificial intelligence and human intelligence. National Science Review, 9(6), nwac068.
Zhuang, Y., Yu, Y., Wang, K., Sun, H., & Zhang, C. (2023). Toolqa: A dataset for LLM question answering with external tools. Advances in Neural Information Processing Systems, 36, 50117-50143.
Zuboff, S. (1988). In the age of the smart machine: The future of work and power. Basic Books, Inc.
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Veda C. Storey is a Distinguished University Professor and the Tull Professor of Computer Information Systems and professor of computer science at the J. Mack Robinson College of Business, Georgia State University. Her research interests are in data management, conceptual modeling, and design science research. She is particularly interested in the assessment of the impact of new technologies on business and society from a data management perspective. Dr. Storey is a member of the AIS College of Senior Scholars and the steering committees of the International Conference of Conceptual Modeling and the Workshop on Information Technologies and Systems. She is a recipient of the Peter P. Chen Award, an ER Fellow, an AIS Fellow, and an INFORMS Fellow.
Wei Thoo Yue is a Professor of Management Information Systems in the Department of Information Systems at City University of Hong Kong. He received his Ph.D. in Management Information Systems from Purdue University. Prior to joining City University of Hong Kong, he was a faculty member at the University of Texas, Dallas. His research interests focus on the economics of information systems. His work has appeared in Management Science, Information Systems Research, MIS Quarterly, Journal of Management Information Systems, Decision Support Systems, and other journals. He currently serves as Senior Editor for Production and Operations Management.
J. Leon Zhao is a Presidential Chair Professor, Director of Center on Blockchain and Intelligent Technology, Co-head of Information Systems and Operations Management, School of Management and Economics, Chinese University of Hong Kong, Shenzhen. He was a chair professor at City University of Hong Kong and Eller Professor at University of Arizona, respectively. He has edited over 20 special issues
for academic journals including MIS Quarterly, Information Systems Research, and Journal of Operations Management. He received an IBM Faculty Award in 2005, and National Chang Jiang Scholar Chair Professorship first at Tsinghua University in 2009 and again at the Chinese University of Hong Kong, Shenzhen in 2022.
Roman Lukyanenko is an associate professor at the McIntire School of Commerce, University of Virginia. His research interests include data management and research methods (validity and artificial intelligence in literature reviews). Roman actively develops ideas, tools, and methods to improve data management and research practices. These solutions received major awards, including INFORMS Design Science Award, Governor General of Canada Gold Medal, Hebert A. Simon Design Science Award. Roman’s research appeared in Nature, MIS Quarterly, Information Systems Research, ACM Computing Surveys. His 2019 paper on quality of crowdsourced data received the Best Paper Award at MIS Quarterly.

  1. J. Leon Zhao
    leonzhao@cuhk.edu.cn
    Roman Lukyanenko
    romanl@virginia.edu
    1 Georgia State University, Atlanta, GA, USA
    City University of Hong Kong, Kowloon, Hong Kong SAR, China
    3 Chinese University of Hong Kong, Shenzhen, China
    4 University of Virginia, Charlottesville, VA, USA
  2. There could be other forms of data, but the basis is textual. Generative AI can “understand” an image because it has seen something similar before. Past images are associated with certain textual data; therefore, new images will also be associated with related textual data.
  3. The authors thank an anonymous reviewer for this positioning.
  4. Transformer-based pre-trained models approach the humanlevel benchmark with General Language Understanding Evaluation (GLUE), which is based on a collection of English language comprehension problems, rapidly. From the GLUE benchmark, a more rigorous SuperGLUE benchmark test was developed, whereby the models rapidly improved and surpassed human-level standards. From OpenAI’s own evaluations, GPT-4 performs exceptionally well on a variety of tests, including reasonings and other examinations (https://cdn. openai.com/papers/gpt-4.pdf).
  5. https://medium.com/mlearning-ai/an-ai-model-that-is-energy-effic ient-is-just-as-important-as-its-purpose-71d17822a183