Reviewer Commitment to Authorial Accountability

In many fields, the scientific paper is the most persistent representation of a research contribution, and the primary interface to the findings for other scientists. In light of its importance to the scientific record, peer reviewers volunteer their expert judgment. Peer review ordinarily proceeds based on an assumption of authorial judgment: that the claims and arguments presented in a submitted paper have been deliberately selected by the authors, and that the authors are prepared to defend them.

Now that generative AI is able to assist researchers throughout the pipeline, from ideation to communication, the terms of peer review change. A reviewer of a paper composed of heavily AI-generated prose is asked to contribute their judgment, but has less basis for assuming that the claims they are evaluating reflect the authors’ judgment. It is up to each reviewer individually to decide if they are willing to contribute their time under these terms.

As submission numbers surge at many venues due to AI-enhanced production, expert human judgment is in greater demand than ever. How we allocate this resource shapes downstream incentives for authors, in turn affecting the future of peer review. When a commodity is scarce, it is rational to be selective about how it is allocated.

We choose to decline to volunteer our judgment as peer reviewers when the representation of a paper we are provided with suggests that the prose is predominantly AI-generated.

We make this choice without claiming that evidence of AI writing is a perfect signal of authors’ involvement or that using evidence of AI writing is a long term solution. Our goal is not to make specific recommendations of methods or evidentiary standards. We recommend only that reviewers have reasonable grounds for the inferences they make and are informed about the limitations of whatever evidence they rely on.

Instead, our emphasis is on the fact that as reviewers, we make individual decisions about where to direct our voluntary labor under uncertainty. As such, we are entitled to reserve our judgment for conditions where there is less ambiguity about author engagement. We make this choice to communicate our desire for high standards on authorial accountability for papers that are sent out for review, and to underscore the importance of ensuring that the scientific record reflects accountable human judgment.

Ultimately, we believe that longer term solutions should not require reviewers to make such inferences nor depend heavily on AI prose detection. The aim of this statement is to incentivize publishing venues to explore mechanisms that can provide stronger assurance of authorial accountability before manuscripts are sent out for review.

This choice should not be read as a statement against generative AI writ large, an insinuation that generative AI has no role to play in scientific writing, nor a suggestion that evidence of AI text alone be used to accept or reject manuscripts. Nor should it be read as condoning reviewers who violate reviewer agreements they have entered into by submitting confidential manuscripts to third-party detection services.