The rise of AI, or LLMs to be specific, has had a significant impact on scientific publishing. From hallucinated references to nonsensical images (that rat penis) and peer reviews devoid of human oversight. More recently, a small number of preprints hosted on arXiv have been discovered to be concealing AI prompts in white text designed to cause chatbots to produce positive reviews. Whilst this undermines the review process1, it also highlights that peer reviewers are using AI tools; whether the research community wants them to or not.
How can AI be used responsibly in peer review?
If peer reviewers are going to use AI, then rather than attempting the futile task of preventing this, we should instead promote responsible and appropriate use. So what is appropriate use of AI in peer review?
Finding reviewers
AI can aid in finding reviewers and matching them to specific manuscripts. This could improve the efficiency of this aspect to the reviewing process, aid editors workloads and reduce the number of researchers rejecting review invitations. Some preprint review services are already utilising AI in this way and this is likely to expand.
Improving grammar in peer review reports
One of the best uses of AI is by non-native English speakers. AI can help reduce the discrimination that such researchers face. This should only be done after the review has been written, by a human, and for grammar only. The review would still need to be checked for clarity afterwards too.
Automated checks
A great use for AI is to automate some tasks to reduce the burden on reviewers and editors; for example in detecting plagiarism. Another great use is to use AI to surface any previous (public) reviews. This could help inform the current reviewers or even be used by the editor to inform their decision.
Summarising human-authored reviews
Transparent peer reviews, posted alongside articles is an important step in improving trust in the scientific process and in individual articles. However, the majority of these reports go unread and are not being utilised in the best way possible. AI could be used to provide summaries of human-authored peer review reports, thereby providing the important context to readers.
Flagging preprints that may need peer review?
A potential new use for such tools would be to flag preprints that may need human scrutiny. This could relieve the stress in the system by avoiding reviewing every output, something that is already unsustainable and failing.
How else could we responsibly use AI in peer review?
Important considerations when using AI
Whenever AI is used, and in whatever capacity, transparency is vital. Use should be declared, including which LLM was used and how exactly it was used. This is important as it provides some degree of confidence that the reviewer or service has used the AI responsibly and checked the results. Indeed, this human-responsibility and oversight is important for all uses of AI; unchecked and unverified content is a large element of AI use that is damaging. Another important consideration when dealing with manuscripts that are not public relates to confidentiality. These manuscripts should not be shared with any third party, which includes uploading them to any LLM model.
We’re currently creating best practices for AI use by preprint servers, preprint review services and authors of both. Want to collaborate with us on this? Please get in touch!
- AI is revealing just how flawed traditional peer review is, supporting the many studies that have investigated issues with peer review.
Photo by Emiliano Vittoriosi on Unsplash
Rippling Ideas 