Abstract
The rapid rise of Large Language Models (LLMs) has introduced both opportunities and ethical challenges to academic publishing. This article, based on an editorial from the INFORMS Journal on Data Science (IJDS), establishes a framework for integrating Generative AI into the research lifecycle. By defining a taxonomy of AI-based "assistants," the authors provide guidance for authors, reviewers, and editors to enhance efficiency while maintaining the core principle of human accountability in scientific integrity.
Introduction
The sudden ubiquity of Generative AI has left the academic world in a state of flux. Researchers are navigating the boundaries of using tools like ChatGPT for drafting papers, while reviewers and editors seek ways to manage the increasing volume and complexity of submissions. The primary challenge lies in the lack of clear normative guidelines. This paper seeks to move beyond the "total restriction vs. total embrace" binary by identifying specific, ethical use cases for AI in the IJDS ecosystem, while simultaneously highlighting the risks of "hallucinations," privacy breaches, and the potential for AI to mask fundamental methodological flaws.
Methodology
As a perspective piece, this study does not rely on traditional laboratory experiments. Instead, the authors utilized a collaborative brainstorming approach, supplemented by queries to GPT-4. The AI was prompted to suggest ways it could assist authors, reviewers, and editors. These suggestions were then critically evaluated against the rigorous standards of the data science field, resulting in a structured taxonomy of AI "assistants" categorized by user role.
A Taxonomy of AI Assistants
The authors categorize LLM utility into several distinct "assistant" roles designed to support different stakeholders in the publishing process:
For Authors
- Writing Assistant: Polishing language and improving clarity.
- Coding Assistant: Debugging and validating research code.
- Creative Assistant: Generating diagrams or visual aids.
- Prereviewing Assistant: Checking for missing references or identifying unclear arguments before submission.
For Reviewers
- Reading Assistant: Summarizing complex sections to grasp the "big picture."
- Verification Assistant: Checking the logical consistency of results.
- Writing Assistant: Helping to structure and refine referee reports.
For Editors
- Integrating Assistant: Synthesizing multiple reviewer reports into a coherent decision.
- Brainstorming Assistant: Identifying potential reviewers or emerging research topics.
Authorship and Accountability
A central tenet of the IJDS framework is the distinction between authorship and assistance. The paper takes a firm stance: an LLM cannot be a co-author because it cannot take legal or ethical responsibility for the work. While using an LLM for language editing is treated similarly to using a professional human editor—as a tool for communication—the underlying scientific contribution must remain human-generated.
Risks and Ethical Considerations
Despite the benefits, the authors warn of several critical risks:
- The "Stochastic Parrot" Risk: LLMs are probabilistic and can produce "hallucinations"—convincing but entirely fake citations, data points, or logical steps.
- Privacy and Data Security: Commercial LLMs often operate as "black boxes." Feeding sensitive, unpublished research or proprietary code into these tools could lead to significant data leaks.
- The "Nice Looking" Flaw: There is a risk that AI-polished manuscripts might look so professional that they distract reviewers from underlying methodological errors.
Conclusion and Future Directions
The IJDS editorial board views these guidelines as a living document. While the journal currently does not require attribution for using LLMs to polish language—provided the methodology is human-generated—transparency remains vital. One significant advantage of these tools is the potential to level the playing field for non-native English speakers, allowing the quality of the data science to shine through. Future work will involve the development of a formal "code of practice" across INFORMS journals and further refinement of AI assistant roles through community experimentation with prompt engineering.
Summary updated