TL;DR

Oxford University Press has updated its AI rules for authors and editors after a survey found three-quarters of researchers do not know what AI use they should declare. Under-reporting is widespread across grant bids, manuscripts and peer review. The survey also shows publishers carry part of the blame: most authors were never asked.

What researchers told OUP

More than 2,600 academics worldwide took part. Although support for disclosure in principle is strong, 74% said the requirements were unclear to them, just 36% keep any record of how they use AI during research, and 44% fear that admitting to it will colour how their work is judged.

Actual behaviour tracks that confusion. Among people who wrote grant applications, 48% had not fully declared AI assistance, often in drafting narrative sections, translation or lay summaries. For journal and book authors the figures were 46% and 39%, and a quarter of peer reviewers had not fully declared AI help either. One respondent admitted deliberately keeping their use “undetectable” because of the uncertainty.

Use is still growing. Almost two-thirds described AI as beneficial, and 68% turn to chatbots on the open web. Trust is thin: only one in five trusts what AI produces, and just over a quarter consider its literature summaries accurate. Attitudes also split by discipline, with 35% of arts and humanities researchers avoiding AI entirely, against 15% in science, technology and medicine.

The new guidance

OUP’s revised policy expects authors of books and journal articles to be open about significant AI use, holds them responsible for checking anything a tool generates, and asks for care when unpublished, copyrighted or sensitive material goes into AI systems.

Looking forward

The most telling figure is about publishers, not researchers. Only 42% of journal authors and 30% of book authors said they were asked to state whether AI had been used. Disclosure rates this low are partly a failure of the question never being put.

For UK universities, where OUP is a department of Oxford itself, the issue reaches beyond journals. Grant bids are where under-disclosure is highest, and institutions are already wary of how AI touches academic work, as Southampton’s decision to drop Turnitin over training-data terms showed. Nearly half of respondents, 45%, described their own institution’s stance on AI as unclear. Clear, consistent rules are likely to come from funders and universities acting together, not publishers alone.