The Evolution of Artificial Intelligence in Peer Review

When ChatGPT launched in late 2022, it raised many questions about the legitimacy of content generated by artificial intelligence (AI). Could the text strung together by AI be trusted? Would authors start using it or be too hesitant and skeptical? How would journals and authorities rein in the unfettered practices of content generation?

Over the last couple of years, many conversations have taken place about the Pandora’s box that is AI. Now that it is widely available, every journal article needs to be verified for its authenticity, and questions about both text and image manipulation seem to crop up more and more. In fact, many obvious anomalies have slipped through the cracks in published papers, which has necessitated the use of AI detection tools and battening down the hatches on AI policy.

What are the current guidelines for AI in scholarly publishing?

Earlier this year, an article by Vasiliki Mollaki published in Research Ethics laid out a comparative summary of policies on the use of AI-based tools by editors, reviewers, and authors followed by large publishers. This revealed that most journals are following COPE guidelines or WAME guidelines. These guidelines for authors boil down to not attributing a large language model (LLM) as an author and disclosing the use and extent of use during submission.

However, while these journals have a structure in place for authors, many large publishers such as Wiley, Sage, and Oxford University Press currently provide little guidance (or no guidance at all) for reviewers. Some publishers, such as Elsevier, use broad language to dissuade reviewers from uploading manuscripts to an AI tool to comply with the author’s data privacy rights. Beyond this measure, because AI has become somewhat ubiquitous in the writing space, the discussion is beginning to shift from “what happens when authors use AI to write their paper?” to “what happens when reviewers use AI to evaluate articles?” Considering Peer Review Week and its importance in scientific discourse, we will take a brief look at the state of AI in peer review.

How can AI be used in the peer review process?

Peer review is a longstanding and important tradition in academia. Often considered the final line of defense between novel research and the public, peer review is, unfortunately, a job that requires a high degree of patience and time. Thus, it has become challenging to secure quality reviews of scientific manuscripts. This has led to some advocates for the use of AI in peer review, whether in part or in whole.

A recent article in Health Affairs Scholar suggests that AI should be used as a screening tool prior to human peer review. In summary, these tools could detect and flag low-quality papers, search for image manipulation, discover potential data fabrication, and more. Such checks would alleviate the pressure on an otherwise overwhelmed peer review system. The article concludes that editors should embrace AI rather than avoid it. The task then becomes to evaluate how strong, accurate, and reliable the technology is and to reassure authors that their work has been fairly considered.

It would seem that AI is a promising approach for weeding out poorly prepared articles, but reviewers would still be writing the actual reports. This application would be akin to human resources departments using algorithms to prescreen candidate resumes or state agencies employing license plate detection to charge drivers on toll roads. The technology would assist editors and reviewers with relatively clerical tasks, not replace the human element.

Could AI replace reviewers?

There is reasonable speculation that, if LLMs can be used to generate content for research papers (with restrictions applied), then there may be some role for the technology to play in fully assessing an author’s paper.

Last year, Donker gave a scathing report about the use of LLMs for peer review. In their report, the LLM generated a review that appeared to be properly balanced but had no critical content about the actual study. It did, though, sufficiently summarize the paper and methodology. Because of this, the report could easily be mistaken for a human-created review by those who have not fully read the manuscript. Such reviews created by AI would be of little use to authors and create more headaches for editors.

This problem is compounded when reviewers use AI tools without notifying the journal or providing disclosure to authors. Peer review, while also acting as a method of academic communication, is a quality control procedure. Journals currently have few options to detect such AI reports, as a dependable AI detection tool has not been developed. Therefore, it is particularly worrying when a quality control procedure requires its own method of quality control.

The obvious flaws in the current landscape of AI generation highlight the need for guidelines to be implemented. Without concrete policies that provide guidance on transparency or penalties for using AI, the integrity of peer review could come into question. Thus far, scholarly journals have relied on third-party organizations to develop guidelines on this matter, so they have either given no direction or outright banned AI-generated peer reviews. It is likely that journals will wait for guidance to come from a higher authority before implementing nuanced policies.

As of right now, authors should rest assured that their research is being reviewed by qualified scholars and not a machine. Whether or not that changes as the technology becomes more sophisticated remains to be seen.