arXiv's AI Author Responsibility Stance Sparks Fresh Debate on Academic Integrity

Thomas G. Dietterich, distinguished professor emeritus at Oregon State University and a foundational figure in machine learning, recently drew attention to a policy that has taken on new significance in the age of large language models. In a post on X, he pointed to arXiv's Code of Conduct, which explicitly states that by signing their name as an author, each author takes full responsibility for all contents, "irrespective of how the contents were generated."

The statement, while not new, represents a clear line in the sand for one of the world's largest preprint repositories. As AI writing tools have become ubiquitous in academic research, questions about where responsibility lies when algorithms contribute substantially to published work have grown increasingly pressing.

The Policy's Origin and Evolution

arXiv's position predates the current wave of generative AI tools. The policy was originally designed to address issues like ghostwriting, honorary authorship, and situations where contributors might not have directly authored every word but still stood behind the work's validity. However, the emergence of LLMs capable of generating coherent, seemingly authoritative text has forced a reexamination of what "responsibility" means in practice.

The policy effectively places the burden on human authors to verify, understand, and stand behind everything in their submissions. This includes results, claims, methodology, and even prose that may have been drafted or substantially modified by AI systems.

Community Reactions

The discussion has generated mixed reactions from the research community. Some view the policy as appropriately strict, arguing that if researchers use AI assistants, they must possess sufficient expertise to validate the output. "If you can't explain and defend every sentence in your paper, you shouldn't be putting your name on it," one researcher commented in response to Dietterich's post.

Others find the policy problematic in its ambiguity. How much AI assistance is acceptable? Where is the line between using AI as a writing tool versus having AI generate the core intellectual contribution? The policy provides no clear answers to these questions.

The Verification Problem

The practical implications are significant. Researchers submitting to arXiv must now grapple with whether their use of AI assistants has crossed a threshold that compromises their ability to take "full responsibility" for their work. Some have adopted conservative approaches, disclosing AI assistance in their submissions or avoiding AI writing tools entirely for paper drafting.

The policy also raises questions about enforcement. arXiv primarily operates on a trust model, and verifying whether authors truly understand and can defend every aspect of their submissions would require substantial resources. The policy functions more as a principle than a mechanism for detection.

Broader Context

This debate sits within a larger conversation about AI's role in knowledge production. Journals and conferences worldwide are grappling with similar questions. Some have begun requiring disclosure of AI assistance, while others have banned AI-generated content entirely. The National Academy of Sciences has discussed updating its guidelines, and several top conferences have added AI disclosure requirements to their submission processes.

What makes arXiv's stance notable is its simplicity. Rather than prescribing specific allowed or prohibited uses of AI, it places the responsibility squarely on authors to ensure the integrity of what they submit. The policy essentially says: the tools you use are your concern, but the final product bears your name and your accountability.

For the research community, this means navigating an evolving landscape where the boundaries of acceptable AI assistance remain contested. Dietterich's reminder of arXiv's existing policy serves as a timely prompt that, regardless of what tools researchers choose to employ, the responsibility for published work ultimately rests with those whose names appear on it.