AI and the Peer-Review Process: Summary of a Discussion at the SIAM Optimization Conference

A special session was held on June 4th, 2026 at SIAM Optimization Conference to discuss the use of artificial intelligence in the peer-review process. The discussion was open to all participants and was well attended. Here we present the key points that came out of this meeting. The purpose is to continue the conversation in the optimization community and inform future journal policies and practices. The discussion highlighted both the potential benefits and significant risks of AI tools, raising questions about confidentiality, correctness, significance, transparency, policy development, and the future role of human reviewers. Many participants emphasized the need for clear and consistent AI policies. Rather than leaving decisions to individual reviewers or editors, several argued that journals and professional societies should establish explicit guidelines governing acceptable AI use. Such policies should clarify what uses are permitted, who is responsible for AI-assisted reviews, how disclosure should occur, and how violations should be handled. It was reported that SIAM journals have adopted a very new policy. This policy essentially prohibits the AI tools unless the authors explicitly allow this. The punishment for violating such policy is severe, but it is so far unclear if detecting a violation is practicable. Other journals are considering their policies. One of the aims of this (and future) discussion is to inform such policies. We now summarize key issues arising around the AI use. Confidentiality: One of the first concerns that was raised was confidentiality. Several participants noted that submitting unpublished manuscripts to commercial AI systems may violate the confidentiality expected in peer review. Unless AI tools are run locally or through systems that guarantee no data retention, the uploaded paper becomes part of a publicly accessible domain. This concern is mitigated when the paper is already posted on arXiv. However, this means that the act of public posting would have a potential effect on the reviewing policies. The solution may be the use of AI tools that do not retain the data or that have some form of a firewall. Who should provide such systems? The journals? If and how AI tools should be used in reviewing. There is a strong agreement than responsibility for a review must remain with the human reviewer. AI may provide suggestions, summaries, or analyses, but reviewers must remain accountable for the final evaluation. Several participants stressed that reviewers should critically assess any AI-generated output rather than simply relying on it. Concerns were raised about the factual accuracy of AI responses and the potential for reviewers to delegate too much intellectual responsibility to automated systems. The distinction between verification of correctness and significance emerged as one of the frameworks. Participants generally agreed than that AI may be useful for checking correctness, identifying technical issues, verifying proofs, or screening for formatting problems. In contrast, judgments about significance, novelty, impact, and future research directions were viewed as inherently human tasks that rely on deep domain expertise and scientific intuition. Several participants suggested decomposing the review process, accordingly, using AI primarily for technical verification while preserving human judgment for evaluating scientific importance. Despite such partitioning of tasks, it is generally agreed that AI systems are not sufficiently reliable to serve as sole arbiters of correctness. Participants noted that proof checking remains difficult, subtle mathematical errors can be missed, and AI-generated assessments often contain mistakes. There is also a significant concern that AI may not “understand” a truly innovative work and flag it as incorrect. The proposed approaches is that AI provides preliminary analysis that is subsequently verified by reviewers. More formally on may consider a review processes in which AI-assisted screening occurs first, followed by more detailed human evaluation. The role of AI in improving readability was viewed more favorably. Participants noted that many papers would benefit from language editing and readability improvements. Some suggested that authors should be encouraged to use AI tools to improve exposition, while reviewers and editors could use AI to identify communication issues. Readability was identified as a third dimension of review, alongside correctness and significance. In contrast, the readability of proofs may suffer if they are not carefully verified by humans. AI tools so far are focused on correctness and are not aimed at making theory elegant and easy to understand. There are concerns this highly valuable property will be overlooked in the future. AI affecting submission rates. The rapid use of AI in research results in a growing burden on editors and reviewers. Multiple participants, including Editors-in-Chief of major optimization journals observed that submission volumes continue to rise, making it increasingly difficult to find reviewers and maintain timely evaluations. This has motivated interest in AI-assisted screening, especially for obvious formatting issues, low-quality submissions, or papers containing substantial AI-generated content. Some journals and conferences have already experimented with such approaches, using automated tools to identify papers that may warrant early rejection or additional scrutiny. However, such screening is likely to become unreliable as submissions adapt to it. Individual comments provided after this summary illustrate these points. A limit on the number of submissions has been mentioned, and it seems that this practice is spreading. While some limits mentioned are very high (36 papers!) a recent post about TMLR policy suggests limits that are more aligned with the “regular” rate of publication in the optimization community. If we decide to rely on primarily (or exclusively) human reviews introducing limits may become necessary. Another major theme was transparency and author choice. Several participants argued that authors should know when AI tools are being used in the review process and should have opportunities to opt out. Similarly, some reviewers expressed discomfort with being required to use commercial AI systems developed by large technology companies. The panel also agreed that the author choice to opt-out should not imply longer review times, although there is some danger of this happening. AI use affecting intellectual advances. The panel also explored broader questions about the future of scientific communication. Several participants expressed concern that excessive reliance on AI could undermine the educational and community-building aspects of peer review. Reviewing is not merely a mechanism for filtering papers; it is also a way for researchers to contribute to the community, learn new ideas, and improve their own understanding. Some worried that offloading too much of the process to machines could diminish these benefits and weaken scientific culture. Related concerns were raised about training future generations of researchers. Current students are growing up in an environment where AI tools are ubiquitous. While many panelists acknowledged that AI will inevitably become part of scientific practice, they emphasized the responsibility of the community to teach students how to use these tools appropriately while preserving the ability to think critically, write clearly, and understand mathematics deeply. In summary, AI is broadly viewed as a tool of potentially valuable use. There is a clear agreement that AI systems should function as tools rather than decision-makers in the review process. Opinions vary on to what extent these tools should be employed without blurring the lines into decision-making. The effect on the future generation of researchers is also unknown and is of concern. Some Individual Comments Madelein Udell: We need to use AI to review papers to handle the ridiculous increase in paper volume now that AI can generate papers. The key is to partition tasks between human editors and reviewers and AI editors and reviewers (and authors) more clearly than ever before. AI can verify correctness of claims (esp using Lean for formal verification). Humans should retain control of taste and significance judgements. This division of labor requires authors to carefully write a list of contributions; if well done, human editors need only consider this list in order to decide how to route a paper. Anonymous: One concern I have with AI in proof checking (for both authors and reviewers) is that it could encourage proofs that are perhaps technically correct, but are not well modularised, humanly meaningful, and amenable to be built on. By analogy with computer science, proofs could gravitate toward becoming brittle machine code that a human cannot meaningfully develop further. Marek Tyburec: One point that I think should have be mentioned today: even if AI can check correctness, it depends on the prompt. In this direction, also the actual paper content can intentionally influence decision by prompts hidden in paper by author. For example, text invisible by eyes (Tex comments, white text in pdf etc....). Anonymous: I feel that the "holy grail" for good reviews is when receiving feedback that fixes subtle errors in creative ways, suggests significant improvements to existing (correct) results, and/or engages with the technical content of the article in a way that only an expert that is passionate about the field would (and certainly beyond a simple correctness check). I am worried that deferring part of this step to AI would dilute the standards we hold the community to, as it relates to technical contributions. We're too quick to jump on the bandwagon as a community, especially given the wealth of evidence on how prolonged engagement with AI affects thinking (and especially diversity of thought). Are we setting the right example for our students? No matter the way forward, authors should have the option to opt out of AI in reviewing, and should not be penalized (e.g., in the form of longer expected review times) for their choice. We send papers to journals to get expert feedback, not get a quick publication by way of AI slop reviews! If reviewing workload is an issue, we should set submission limits and be transparent about review times. Nicolas Boumal: A point was made in the discussion about editorial offices using tools to detect AI-generated papers, for automated desk rejection. I believe (a) we can't, (b) we shouldn't and (c) we don't need to do this. (a) It's not possible to detect AI-generated text reliably enough to then act on it with drastic measures such as a desk reject, let alone a journal ban as was floated at some point. Detecting hallucinated references is doable, but this is a cat-and-mouse / stop-gap solution: models will improve on this. (b) I don't think we should automatically reject papers that contain a large fraction of AI-generated text, because nuances matter a great deal here. To elaborate on just one of these: Most researchers are not native English speakers. I can imagine that some authors may soon decide to write papers in their native language, and then have this translated to English. De facto, their entire paper would have been written by AI (and likely would be flagged as such by good detectors), yet in that scenario all of the ideas and the narrative are the author's original contribution. I don't want to block this use case. (c) We already reject papers which we deem uninteresting. If, as editors, we receive a paper which we find pointless, we can reject it on those grounds. Even if we surmise that it was AI-generated, there is no need to include this in the rationale for the rejection, as it would be hard to defend with certainty anyway. Besides, if an AI-generated paper today would have been considered a valuable contribution if done by a human today, do we not want to add it to the published record? I would say that we do, with proper disclosure. In time, goal posts will shift: the bar for publication has always dynamically adapted to match what is deemed interesting / difficult to achieve, in light of all available tools. I understand the pressure that journals are under. As Daniel Kuhn and another editor in chief eloquently stated, their job of desk-rejecting subpar papers has gotten a lot harder, and we need to develop processes to reduce that burden. I do not have a good proposal at this time, but I did want to comment that I do not believe AI-detector-based automatic-desk-rejectors are the way forward. I could imagine that they can serve as an internal flagging system to assist editors, but not to automate decisions. Anonymous: I am not objecting to the use of AI in either writing, review or even any activity. However, I think we should always keep two principles in our mind whenever we use AI tools to perform a certain task: be responsible for the outcome. For e.g., ensure that the proofs are correct to one’s best knowledge for deciding to accept an article, and this should be the requirements for either the reviewers or the authors. Maybe in the future, with the help of rigorous AI tools (such as format language), this process can be significantly sped up or even automated. However, before that day comes, mathematics related professionals should not wave hands and trust AI entirely (LLM is far from perfect). Second, the reviewer need to make sure that the decision made is to their best judgement and supported by valid reasons. The comments should represent that of their own. always think about what makes us human non-irreplaceable when performing the task. If the review process can be purely performed by AI, then soon human will not be involved in the review process. So does the writing or any tasks. Reviewer and editors, when using AI tools, need to think about this. If AI can do my job for me and even a better one, why am I still needed in this? What should be human’s contribution? Can human-AI work pattern outperform each alone?