AI and Peer Review 2025
Insights from the global reviewer community
Results
Key findings
- Views of the future impact of generative AI on peer review remain polarized among physical sciences researchers.
- Women tend to feel less positive about the future impact of generative AI on peer review compared with men.
- More junior researchers tend to be more positive about the impact of generative AI compared with more senior researchers.
57%
of respondents said they would be unhappy if a reviewer used generative AI to write a peer review report on a manuscript that they had co-authored
42%
of respondents said they would be unhappy if a reviewer used generative AI to augment a peer review report
42%
of respondents thought that they could accurately detect an AI-written peer review report on a manuscript they had co-authored
32%
of respondents admit using AI tools to support them with the peer review process
Perceived future impact of AI on peer review
We asked respondents: “In your opinion, what impact will open-source generative AI tools, such as ChatGPT, have on the peer review process?”
37%
(128 of 343) believed that generative AI would have a negative or very negative impact on the peer review process
22%
(77 of 343) said it would have a neutral or no impact
41%
(138 of 343) said it would have a positive or very positive impact
Compared with 2024, there were substantially fewer respondents who were neutral about the impact of AI.
36%
of respondents were neutral in the 2024 survey
2%
rise in respondents predicting a negative impact of AI on peer review since 2024
12%
rise in respondents predicting a positive impact of AI on peer review since 2024
Gender analysis
When the responses were analyzed by gender, roughly the same proportion of men and women were neutral towards AI in peer review, but women tended to have a less positive and more negative view of generative AI than their male counterparts.
Perceptions of AI impacts by career stage
When the responses to this question were analysed by career stage, with career stages grouped into more junior (undergraduate or Master’s degree, PhD, postdoctoral researcher) and more senior (early-career researchers ECR), faculty, associate professors or higher), with the groups ‘I work in industry’ and NA removed, the results show that more junior researchers tend to have a more positive view of the impact of AI on peer review than their more senior counterparts.
In total, 48% of junior researchers thought that AI would have a positive impact on peer review compared with 34% of more senior researchers.
Free-text responses
As in our 2024 survey, we followed this question up with an optional free-text response. 233 respondents added something to this field, and many of the responses were lengthy and nuanced.
The table below includes a selection of negative, positive and neutral comments received. To view full table of responses, click here.
| Negative perception of the impact of AI on peer review | Neutral/no impact | Positive perception of the impact of AI on peer review |
|---|---|---|
| “The definition of peer review is that it is done by peers. AI is not an academic peer. AI can only draw from what is already in the LLM. Research should be making a new contribution. These 2 things are in conflict” | “AI can accelerate peer review by summarizing papers, checking references, but it might also reduce the diversity of perspectives. AI could act as a filter or first-pass reviewer.” | “It can perfectly answer questions and is a tool that can drive technological progress.” |
| “All papers prepared with the help of AI are very eloquent and immensely stupid. I consider AI as the profanation of true science. The mass media methods like AI are very dangerous infection able to keel any rational human thinking.” | “The impact of AI on peer review is neutral when reviewers engage sincerely with the manuscript and judiciously leverage AI as a supportive tool.” | “It will make the process faster, it will make it easier to understand the key concepts of an article saving valuable time and allowing for better focusing on what is novel and important to be checked thoroughly.” |
One respondent admitted to using AI to write their response to the question:
‘Generative AI is inevitably becoming a part of the academic workflow, and its integration into the peer review process is not only foreseeable but necessary. Dismissing its use would be a missed opportunity to enhance efficiency, consistency, and access to knowledge. However, it is essential that its use be approached with responsibility and critical oversight. AI can make mistakes, misinterpret context or generate plausible sounding but incorrect content. Therefore, human reviewers must remain central to the process, verifying and contextualizing AI-generated suggestions or evaluations. When used wisely, AI has the potential to assist in identifying inconsistencies, improving clarity and even flagging ethical issues—thereby complementing human judgment rather than replacing it. This response was written by AI based on a personal reflection that stated: ‘AI is coming whether we like it or not. Not using it would be simply absurd – as long as we do so responsibly, since it also makes mistakes. One must always verify what it says, but without a doubt, it is a tool that will change the world.’
Previous use of AI tools in the peer review process
We asked respondents: “Have you ever used any AI tools to support you with the peer review process?”. 32% of respondents (113 of 348) answered yes, and the remainder answered no. We then asked: “Which tools have you used to help with your peer review work?” The most popular answer by quite a large margin was ChatGPT, followed by Deepseek, and Gemini.
However, it should be noted that 48% of respondents selected more than one answer for this question.
The ‘Other’ AI models that were mentioned are (in no particular order): Paperpal, Scite, Grok, Connected papers, Turnitin and Notebook LM.
Then we asked: “Which of these ways have you used AI tools to help with your peer review work?”. 113 individuals responded to this question, which had multiple choice options and an ‘Other’ option with free-text response.
The most popular answer was “Putting a completed review into an AI tool to improve flow and grammar”. Again, it should be noted that 53% of respondents selected more than one answer for this question.
| Response | Count |
|---|---|
| Putting a completed review into an AI tool to improve flow and grammar | 73 |
| Using AI tools to digest or summarize an article under review | 46 |
| Using AI tools to translate a review from another language into English | 30 |
| Other (free-text comments) | 24 |
| Using AI tools to turn a simple review or bulleted list into a longer review | 15 |
| Uploading part of an article (such as title or abstract) to an AI tool and asking it to create a review on your behalf | 6 |
| Uploading a full article to an AI tool and asking it to create a peer review report on your behalf | 6 |
While some of the free-text comments mirrored the survey response options, there were two additional themes captured in the comments; reviewers using AI to clarify unclear sections or unfamiliar topics in the manuscript, and using AI to assist with the review of the literature/references.
Other stand-alone comments mentioned using AI to: create a report template to fill out; detecting AI use in the manuscript; testing models in the manuscript; and doing an initial assessment of the manuscript.
Perspectives on AI in peer review
We asked respondents a series of questions regarding how they felt about different aspects of generative AI in peer review, each of which had the following response options: Strongly disagree, disagree, Neutral, Agree, Strongly agree.
In response to the question: “I would be unhappy if peer reviewers used AI to write their peer review reports on a manuscript that I had co-authored”, 57% of respondents (195 of 345) either agreed or strongly agreed, with 24% responding that they felt neutral and only 19% either disagreeing or strongly disagreeing.
In response to the question: “I would be unhappy if peer reviewers used AI to augment their peer review reports on a manuscript that I had co-authored”, the proportion of respondents saying that they
either agreed or strongly agreed went down, to 42%, with 31% being neutral and 28% either disagreeing or strongly disagreeing.
In response to: “I feel that AI tools could add value to peer review reports and help to improve manuscripts”, 24% disagreed or strongly disagreed, 27% were neutral and 49% agreed or strongly agreed.
We then asked respondents to rate how strongly they agree with the statement: “I feel that I would be able to accurately identify AI-authored or AI-augmented peer review reports if they were submitted on a manuscript I had co-authored”.
42% agreed or strongly agreed; 36% were neutral, and 22% disagreed or strongly disagreed. With a free-text response, we asked respondents, “What (if any) do you think are the hallmarks of AI-generated peer review reports?”.
223 respondents chose to write something in the free-text box.
A small selection of the responses is given in the table below. Recurring themes included AI-generated reports being highly structured, being well-written from a technical perspective but lacking depth, and containing equal amounts of positive and negative comments regardless of the manuscript.
| Selected comments on the hallmarks of AI-generated peer review reports |
| “Hallmarks of AI-generated peer review reports may include overly generic language, lack of specific critique, uniform tone, and repetition of phrases without deep engagement with the manuscript’s content.” |
| “Superficial and heavily structured reports lacking substantial technical depth.” |
| “Circular reasoning and nonsensical explanations.” |
| “Human reviewers often draw on their own research experience or give contextual advice; AI cannot do this meaningfully.” |
| “Generic, overly flowery language; mistaking words frequently based on alternate definitions that no human with knowledge in the field would do.” |
| “I don’t think there are many. AI-generated reports will likely become indistinguishable from actual reports in the near future. This, however, does not make them more reliable.” |
| “The language style often lacks natural flow, and AI does not always fully catch the underlying physics. It may fail to infer certain points that are implied but not explicitly stated in the article, something an expert reviewer would typically recognize with ease. At the moment, I think, AI lacks that level of domain-specific understanding.” |
| “I can’t characterize it exactly, but I can identify it. A machine is not a human. A machine cannot think.” |
| “I think the biggest giveaway of AI-generated peer reviews is that sometimes the comments they make/generate are blatantly wrong.” |
| “As far as I’ve seen, the tell-tale sign of an AI-generated review is the lack of any relevant, subject-specific critiques: Obviously, AI or not, a review is bound to have a few comments dedicated to more surface-level issues, e.g. A typo in an equation, the lack of clarity in an explanation, under-detailing something within a computational procedure, etc. The problem with AI reviews is that they essentially stop there: They do critique the surface-level issue of ‘Why was this model chosen?’, but fail to have the expertise to ask ‘Why is X model chosen over Y? I would assume it could make a difference, so have you run tests to compare the two?’. While the former questions does slightly improve the quality of the paper, the latter actually might reveal a major weakness/oddity of the paper that the peer-review process should actually catch but LLMs don’t.” |
With a free-text response, we asked respondents: “What (if any) do you think are the ethical issues surrounding use of AI in peer review?”.
239 people chose to respond to the question. A selection of the responses is below.
Recurring themes included a lack of accountability for AI outputs, the risk of fabrications or hallucinations, issues around confidentiality of manuscripts under review, issues around honesty and transparency, and built-in biases of LLMs trained on biased data.
There were also several responses stating that they did not see any potential ethical issues regarding AI-generated reviewer reports.
| Selected comments on the ethical issues surrounding use of AI in peer review |
| The theft of the corpus of data used to train the AI models, the replacement of human labour, the wasteful energy usage. |
| The main ethical issue is the transfer of responsibility over knowledge from a human intelligence to a non-biological intelligence with unknown administrators or proprietaries. |
| If the author wants the insights of AI, they can easily do it on their own. When they’re asking for a peer review, they expect an honest review from the researcher. |
| Is it still a ‘peer’ review if it was written by an AI? Who is responsible if an article gets incorrectly accepted or rejected based on a review written by an AI? |
| The primary ethical concerns include confidentiality breaches when unpublished manuscripts are uploaded to AI platforms, especially if data is stored or processed externally. There is also the risk of bias propagation, lack of transparency about AI involvement and potential over-reliance on tools that cannot replicate expert judgment. Using AI without disclosing it undermines the integrity and trust in the peer review process. |
| I don’t see any kind of ethical issues in AI peer review; I just feel there might be less bias. |
| Uploading manuscripts to public AI tools (like ChatGPT or Gemini) risks exposing sensitive, unpublished research. AI systems lack domain expertise and may produce plausible sounding but factually incorrect reviews. |
| As long as confidentiality is maintained prior to acceptance I see no significant issue – when used it should be noted for the sake of transparency. |
| Poor work might be published after going through a poor peer review process. People can get credit for doing a review without even reading the manuscript. Public trust in scientific outputs would be at risk. |
| The ethical risks of AI in peer review stem from its potential to erode the trust, fairness and human agency that define scholarly evaluation. |
The final field on the survey was a free-text form: “If you have any more thoughts or comments about the use of AI tools in peer review not covered by the questions above, please add them here”.
113 respondents chose to add text to this field. The responses covered a wide range of topics and viewpoints. A sample of comments is included in the table below.
| Responses |
| Tools that help correct spelling and grammar are standard and have been for a while. Anything where the intellectual work is outsourced is unacceptable. The problem is that peer review relies on voluntary work by academics and so is vulnerable to AI to help cope with the workload, especially as the rate of submission of papers has increased, again with the impact of AI. I am very concerned about the tendency to quantity over quality. |
| A useful tool that accelerates and facilitates the reviewer’s work. Obviously, the final evaluation of the work must remain with the human reviewer. |
| My personal preference for the use of generative AI and LLMs is to support writing or content I have already generated myself and to help me to ask other questions that I might not have seen before (…). It is not necessarily unethical to use an LLM to assist with easily automated tasks such as memory or grammar/spelling or even to support human tasks such as creation, planning and context analysis. But that there is the crux of the matter, AI is fast approaching a point where it is replacing distinctly human activity itself. |
| The core issue is that AI cannot replicate the depth of insight and critical engagement a human reviewer provides. |
| While AI has potential to assist with grammar or structure, its role in evaluating scientific rigor, methodology, or novelty should remain secondary to human expertise. Clear policies and training are essential to guide ethical and responsible use. The future of peer review may benefit from AI support, but not AI substitution. |
| If AI is to be extensively used in place of human referees, innovative ideas would face a challenging time. |
| I will not submit papers to or undertake peer review for any journal that allows LLMs to be used anywhere in the peer review process. |
| AI in peer review also needs transparency. Plus, there is potential for AI-human teamwork: AI could handle basic checks, freeing humans for deep critiques. But we must keep balancing their roles to avoid new scholarly blind spots as AI evolves. |
| I am 100% against AI tools in ALL stages: (1) writing a paper (2) proofreading and editing a paper, (3) reviewing a paper, after submission to a journal. |
| The old phrase “use it or lose it” comes to mind. If reviewers rely on AI very much, their own ability to make a critical assessment of the work will decline. |