Source-linked AI summary
Generative AI
Stefan Feuerriegel, Jochen Hartmann, Christian Janiesch, Patrick Zschech
TL;DR
Generative AI raises important questions for information systems and BISE because current systems have limitations, including bias and fairness concerns. The paper conceptualizes generative AI across model, system, application, and socio-technical levels, reviews its implications, and develops a BISE research agenda. It concludes that generative AI has substantial implications for BISE practitioners and scholars and offers impactful research directions.
Problem
The paper addresses the need to understand generative AI in information systems while accounting for current limitations such as bias and fairness concerns.
Method
The paper conceptualizes generative AI through model-, system-, application-, and socio-technical-level views and uses these perspectives to develop a BISE research agenda.
Results
The paper describes limitations of current generative AI and identifies opportunities, challenges, and research directions relevant to BISE.
Takeaways & Limitations
Generative AI has substantial implications for BISE practitioners and scholars across the interdisciplinary research community.
Takeaways & Limitations
Training deep learning models on biased data can amplify societal biases, making bias and fairness a limitation of current generative AI.
Abstract
from arXiv · showhide
The term "generative AI" refers to computational techniques that are capable of generating seemingly new, meaningful content such as text, images, or audio from training data. The widespread diffusion of this technology with examples such as Dall-E 2, GPT-4, and Copilot is currently revolutionizing the way we work and communicate with each other. In this article, we provide a conceptualization of generative AI as an entity in socio-technical systems and provide examples of models, systems, and applications. Based on that, we introduce limitations of current generative AI and provide an agenda for Business & Information Systems Engineering (BISE) research. Different from previous works, we focus on generative AI in the context of information systems, and, to this end, we discuss several opportunities and challenges that are unique to the BISE community and make suggestions for impactful directions for BISE research.
1 Introduction
Generative AI generates seemingly new, meaningful text, images, or audio from training data and is changing work and communication. The article examines it in information systems, identifies limitations, and proposes impactful directions for BISE research.
- Generative AI comprises computational techniques that generate seemingly new, meaningful content, including text, images, and audio, from training data.
- Recent systems such as Dall-E 2, GPT-4, and Copilot are transforming how people work and communicate.
- Industry reports suggest generative AI could increase global GDP by 7% and replace 300 million knowledge-worker jobs.
- The article conceptualizes generative AI as an entity in socio-technical systems and provides examples of models, systems, and applications.
- The article introduces limitations of current generative AI and suggests impactful directions for BISE research.
- Unlike work focused on particular methods or applications, the article studies generative AI in information systems and identifies opportunities and challenges unique to BISE.
2 Conceptualization
The paper conceptualizes generative AI across model, system, and application levels, while distinguishing generative modeling from discriminative modeling. It also describes multimodal models, varied training procedures, and socio-technical human–AI co-creation.
- Mathematical Principles: Generative modeling infers data distributions to produce new synthetic samples, unlike discriminative modeling, which separates data into classes using decision boundaries.The paper gives P(X,Y) and P(Y) as examples of distributions that generative models may infer.
- Conceptualization: Generative AI is structured at model, system, and application levels, spanning underlying models, interaction infrastructure, and practical use cases such as SEO and code generation.The system includes data processing and user-interface components, while applications address real-world problems.
- Model-Level View: Models are organized by output modality and may be unimodal or multimodal, with examples covering text, images, audio, music, and code.Unimodal models use the same input and output type; multimodal models can combine different input sources and output forms.
- Mathematical Principles: Advanced models often combine generative, discriminative, and reinforcement-learning mechanisms rather than relying on a single modeling principle.The paper cites GPT-family pre-training and fine-tuning, and ChatGPT’s combination of generative modeling, discriminative modeling, and reinforcement learning.
- Training Procedures: Generative AI training procedures vary substantially, including GANs’ competing synthesis-and-detection objectives and ChatGPT-based systems’ reinforcement learning from human feedback.RLHF uses demonstrations, user rankings of outputs, and policy learning through reinforcement learning.
- Socio-Technical View: At the socio-technical level, generative AI supports human–AI co-creation, while perceived artificial minds may influence usage intensity and collaboration.The paper connects co-creation with collective intelligence and proposes that understanding perceived AI states and rationale may alleviate collaboration concerns.
3 Limitations of Current Generative AI
Current generative AI is constrained by technical, social, legal, and environmental limitations that affect reliability, fairness, accountability, and sustainability. These limitations arise from probabilistic generation, data and model biases, intellectual-property risks, and resource-intensive development and operation.
- Technical limitations: Because models generate the most probable rather than necessarily correct response, outputs can resemble authentic content while presenting misinformation or deceiving users.
- Technical limitations: Hallucinations produce semantically or syntactically plausible content that is actually nonsensical or incorrect.
- Technical limitations: Generative AI outputs are probabilistic and typically lack transparent factual grounding, making correctness dependent on training data, learning processes, and user trust.Closed-source commercial systems further restrict model tuning and retraining.
- Societal limitations: Biases in training data, alignment, model engineering, and reinforcement learning can amplify stereotypes, toxic language, and unfair outputs.System- and application-level mitigation mechanisms can increase diversity, but fair AI remains an open research question.
- Legal limitations: Generative AI can violate copyright by reproducing protected works, generating similar logos, or producing derivative works with unclear intellectual-property ownership.
- Environmental limitations: Large-scale generative AI models consume substantial electricity and can produce substantial carbon emissions during training and operation.Training GPT-3 was estimated to produce 552 t CO2, equivalent to the annual emissions of several dozen households.
4 Implications and Future Directions for the BISE Community
The paper outlines BISE research opportunities arising from generative AI across business processes, knowledge management, modeling, interaction, communication, marketing, education, and societal applications. It emphasizes interdisciplinary, socio-technical inquiry into applications, organizational consequences, governance, and rigorous causal evidence.
- Research agenda: BISE research should examine generative AI as an application-oriented, socio-technical phenomenon and organize questions across the journal’s interdisciplinary departments.
- Business Process Management: Generative AI can support BPM design, process extraction, implementation, simulation, predictive monitoring, and execution across the process lifecycle.Prompt engineering may support process extraction from text without further fine-tuning.
- Business Process Management: Applications include generating process descriptions, supporting process redesign and innovation, automating routine rules, and enabling intelligentized legacy software.
- Knowledge and process guidance: Enterprise-wide language models could improve retrieval from heterogeneous sources and enable new process-guidance systems for knowledge-intensive work.Relevant sources include manuals, handbooks, e-mails, wikis, and job descriptions.
- Value creation and knowledge management: Generative AI creates opportunities for knowledge discovery, automated content sharing, personalized delivery, new business ideas, product and service innovations, and new business models.
- Organizations and strategy: Future BISE studies should investigate work patterns, competition, competitive advantage, API management, and how existing theories and frameworks must be contextualized or extended.The paper calls for quantification through rigorous causal evidence given the velocity of AI research.
- Enterprise Modeling and Engineering: Generative AI can assist enterprise modeling by creating and updating models at different abstraction levels, generating multimodal representations, and producing conceptual models from text.Reported examples include ER, BPMN, UML, and Heraklit models generated from textual descriptions with very high to perfect accuracy.
- Interaction and applications: Natural-language interfaces, communication tools, marketing systems, recommender systems, education, harmful-content detection, and crowdsourcing are further application and research areas.Educational governance should address appropriate reliance, output verification, and prompt engineering.
5 Conclusion
The article positions generative AI as a consequential topic for BISE by conceptualizing it across model, system, application, and socio-technical levels. It also identifies current limitations and develops a research agenda focused on BISE affordances and implications.
- Generative AI creates new content such as text, images, or audio and increasingly produces outputs difficult to distinguish from human craftsmanship.
- Its capabilities may transform creativity-, innovation-, and knowledge-processing domains by enabling applications previously impossible or impractical to automate.
- The article conceptualizes generative AI through model-, system-, application-level, and socio-technical perspectives.
- The authors describe limitations of current generative AI and provide an impactful research agenda for BISE.
- The BISE perspective highlights generative AI’s manifold affordances and its substantial implications for practitioners and scholars.