Medical reviewers know AI—but 70% in a global survey had not used it for peer review
An international survey found high familiarity with chatbots alongside limited peer-review use, scarce training and widespread concern about bias. Its 1.9% response rate sharply limits generalisation.
The short version
- What happened
- Of 1,077 eligible respondents answering the main survey, 70.3% said they had never used an AI chatbot for peer review, despite 86.2% being familiar with the tools.
- Who it affects
- Authors, patients and readers rely on confidential, technically competent review; unclear AI use can expose unpublished material or add fluent but incorrect criticism without anyone recognising the source.
- What remains unproven
- 1.9% response rate among delivered invitations
Keep in mind
1.9% response rate among delivered invitations
Self-reported attitudes and behaviour
English-only survey
Corresponding authors were screened for reviewer eligibility after responding
Cross-sectional design cannot measure change or causal effects
No direct test of peer-review quality
Source analysis · 1 linked source. Primary analysed source date: 26 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
Familiarity has moved faster than peer-review practice
A peer-reviewed international survey found a practical gap between knowing about AI chatbots and using them in medical-journal peer review. Among 1,077 eligible respondents who answered the main questions, 86.2% were familiar with chatbots and 87.7% had used ChatGPT generally, yet 70.3% said they had never used an AI chatbot for peer review.
The more immediate issue for reviewers and editors is institutional support. Of 1,067 respondents, 69.5% reported no institutional training; 60.7% were interested in training. More than half said their institution had no policy for chatbot use in review, while concern about algorithmic bias reached 80.3%. The survey measured perceptions and self-reported behaviour—not review quality, accuracy or misconduct.[1]
A large respondent group came from a very low response rate
The team drew corresponding-author contacts from MEDLINE records, sent 72,851 invitations and recorded 6,172 bounces. It received 1,260 responses from 66,679 delivered emails: 1.9%, or 3.7% among emails recorded as opened. Respondents worked across all six WHO regions, but the Americas and Europe supplied 73.6% of the main sample and the English-only questionnaire excluded many researchers.
For journals, the finding supports clear confidentiality, disclosure and verification rules before broader adoption; it does not show that AI improves peer review. Confidence would rise with representative recruitment, observed review behaviour and blinded comparisons of human-only and AI-assisted reports for error detection, bias, time, confidentiality breaches and author benefit. The study was preregistered, unfunded and reported no competing interests.[1]
The distinction that matters
What the survey can—and cannot—establish
The broad invitation list produced useful descriptive evidence, but not a population estimate of all medical reviewers.
Invited
66,679 delivered emails
1,260 responses; 1.9% response rate
Main sample
1,077 answered core questions
Report familiarity, use, training and concerns
Not tested
No review-performance comparison
Accuracy and quality effects remain unknown