ChatGPT can provide researchers with useful feedback on their papers, suggesting it could supplement the human peer review process that helps scientific journals decide which studies to publish. But others are sceptical that ChatGPT could play a role in peer review.
Peer review is a critical component of scientific publishing. However, many journals and research conferences are struggling to recruit enough human peer reviewers – who typically volunteer their time for free – to evaluate the growing number of submitted papers.
“It’s getting harder and harder for researchers to get high-quality feedback from reviewers,” says at Stanford University in California. “That motivated us to see if we can use large language models like ChatGPT to provide feedback to these researchers.”
Advertisement
Zou and his colleagues first asked OpenAI’s ChatGPT – specifically the chatbot version based on the AI model GPT-4 – to provide constructive feedback on more than 3000 manuscripts that were submitted and accepted to the prestigious after undergoing standard peer review. They also asked the chatbot to review more than 1700 papers from the (ICLR), one of the world’s largest machine learning conferences. They then compared the content of the AI’s critique with comments from human reviewers who had assessed the same papers.
More than half of ChatGPT’s feedback on the Nature papers made points that were also raised by at least one of the scientists who peer-reviewed the manuscripts. For the ICLR conference papers, more than 77 per cent of ChatGPT’s comments highlighted points that were also raised by at least one human reviewer.
Separately, the study also used ChatGPT to generate feedback on several hundred papers that have yet to be peer-reviewed, but that were recently uploaded to preprint servers. Zou and his colleagues surveyed 308 of the authors – all researchers in the fields of AI and computational biology – for their impressions of the critiques.
Although they made no direct comparison with any human reviewers’ comments, more than 82 per cent of the researchers described ChatGPT’s feedback as generally more beneficial than some of the feedback they had received in the past.
But the study’s lack of detail about cases when AI-generated feedback differed from human reviews “disguises the extent to which AI-generated reviews can be extremely biased or erroneous,” says at Northwestern University in Illinois.
Zou says some of the researchers they surveyed described ChatGPT’s feedback as too vague and lacking technical detail. He thinks AI could provide researchers with some initial comments on their papers. But he cautions against peer reviewers relying on ChatGPT to do their work. “We still do need to have high-quality human reviewers,” he says.
Zou and his colleagues have posted their work to the web – but it has yet to be peer-reviewed and published by a scientific journal.
arXiv
