AI-assisted peer review platforms and tools

Last updated: 2026-10-20

Synthesis

AI in peer review is the most contested adjacent space to scriptorium. The work falls into four buckets: (1) statistical and methodological check tools that pre-screen manuscripts for reporting completeness (StatReviewer, Penelope); (2) citation-intent and integrity tools that operate post-publication or at submission (scite.ai, Springer Nature’s irrelevant-reference checker, IOP Publishing’s duplicate-review checker); (3) LLM-as-reviewer systems that simulate or generate critical reviews (Stanford PaperReview.ai, AgentReview, OpenReviewer, ReviewerToo, Liang et al.’s GPT-4 pipeline, the AAAI-26 AI review pilot); and (4) platform-level innovation in how peer review is organized (OpenReview, ResearchHub).

The strongest empirical signal in the space is still Liang et al.’s study: GPT-4 feedback on 3,096 Nature-family papers and 1,709 ICLR papers overlaps with individual human reviewers at roughly the same rate that two human reviewers overlap with each other (~30% Nature, ~39% ICLR) [6]. Three 2026 results extend rather than replicate it. A randomized study at ICLR 2025 (Thakkar et al., Nature Machine Intelligence 2026) gave LLM feedback to reviewers on more than 20,000 randomly selected reviews; 27% of reviewers who received feedback updated their reviews, and updated reviews were longer and rated more informative [7]. AAAI-26 gave every one of 22,977 main-track papers a clearly identified AI review, and surveyed authors and committee members preferred the AI reviews to human reviews on dimensions such as technical accuracy and research suggestions [18]. And an expert-annotation study of 2,960 criticisms on 82 Nature-family papers found the best AI reviewer’s criticisms scored above each paper’s top-rated human reviewer on a composite of correctness, significance, and evidence (60.0% vs. 48.2%), while AI reviewers overlapped with each other far more than humans did (21% vs. 3%) and showed recurring weaknesses humans did not; its authors position current AI reviewers “as complements to, not substitutes for, human reviewers” [17]. This is the strongest evidence yet that LLM critique is non-trivially useful; it does not justify replacing human peer review — only augmenting it.

Reviewer-side policy moved in 2026 from near-blanket bans toward risk-tiered, disclosed use. NIH still prohibits peer reviewers from using generative AI to analyse or formulate critiques of grant applications [21]. Elsevier’s journal policy (updated June 2026) lets reviewers use private AI tools only in a supportive capacity (e.g., language and structure of the report, background literature search), requires disclosure, and says reviewers should not upload a submitted manuscript or any part of it into an AI tool (its FAQ says “must not”) [22]. Springer Nature’s reviewer guidance, on an undated policy page, sorts reviewer AI use into green (e.g., clarity, tone, and structure of comments; generating checklists), amber (e.g., “exploring potential limitations; stress-testing reviewer reasoning”), and red (uploading manuscripts to public or unsecured tools, generating full review reports, producing accept/reject recommendations, delegating critique) tiers [23]. Its 10 September 2026 policy update describes the overall shift to judging AI use by how it was used, its impact, and the level of risk, and says delegating peer review to an AI tool remains not permitted [24]; the release does not itself describe the colour tiers, so whether the tier page was issued with that update is not established. ML conferences require disclosure and accountability (ICLR 2026) or sanction only a conference-provided, logged LLM interface inside a randomized experiment (NeurIPS 2026) [25, 28]. Integrity problems are now documented at scale: “review mills” that reuse boilerplate reports and coerce citations [29, 30], nearly 2,500 near-duplicate reports flagged among around half a million at one publisher [15], 21% of ICLR 2026 reviews reported as fully AI-generated by an AI-detection analysis [26], and manuscripts that hide prompt-injection instructions aimed at AI reviewers [19].

For scriptorium, the key positioning is: reviewer-simulation is an author-side critique tool that helps a manuscript get ready for human peer review. It is explicitly not a substitute for peer review and should never be marketed as one.

Landscape

Methodological / statistical pre-screening

StatReviewer (Houle & DeVoss). Pre-LLM, rule-based tool for statistical and reporting integrity, created by Timothy Houle and Chadwick DeVoss [1]. Per the 2017 Aries Systems announcement, it runs “thousands of algorithmic tests” covering “obvious numerical errors, quality of reporting (e.g., defined descriptive statistics), appropriate statistical tests (e.g., t-test for skewed data), style (e.g., precision of decimal places) and methodological reporting,” and produces a checklist of statistical issues for editors; Aries planned to integrate it into Editorial Manager in 2018 [1]. The tool’s positioning has remained “pre-review screening to augment human reviewers,” and it predates the LLM era by years — a useful reminder that rule-based tooling can do real work without generation. (An earlier version of this note also listed specific CONSORT/STROBE/ICMJE checks, IMRAD detection, and the creators’ institutional affiliations; these were not re-verified against a source in the 2026-10 refresh and have been removed.)

Penelope.ai. Covered in ai-writing-tools-survey; the closest production example of inspectable, journal-side compliance checks. 30+ checks per manuscript, customizable per journal, linked back to manuscript text.

Citation-intent and integrity tools

scite.ai (Nicholson et al.). Published in Quantitative Science Studies (MIT Press) [2]. Uses deep learning to classify citation context as supporting, contrasting (called “disputing” in the paper), or mentioning. The preprint version reports a database built from over 23 million full-text articles with more than 800 million classified citation statements, and an estimated average distribution of 92.6% mentioning, 6.5% supporting, 0.8% disputing [31]. (An earlier version of this note gave 25M+ articles and 880M+ statements; those figures could not be checked against the paywalled published version in this refresh.)

Critically, the developers’ own paper and the independent evaluation point in different directions. The supporting/contrasting distinction is the value proposition, but an independent evaluation in Hypothesis (Bakker, Theis-Mahon & Brown 2023) found that only 98 of 324 sampled citations were classified by scite, that scite labelled 96 of those as mentioning where human assessors found 42 supporting and 17 contrasting, and that F-measures ranged from 0.0 to 0.58 [3]. scite’s own pipeline reports ~70% citation-context-to-DOI association for raw PDFs and ~95% for structured JATS XML [31]. The product is widely deployed in university libraries; the literature on its accuracy is unsettled.

Publisher integrity tools (2024–2026). Springer Nature launched an “irrelevant reference checker” for submitted manuscripts in April 2025, flagging submissions to its Research Integrity Group for manual review, alongside earlier tools for AI-generated fake content (Geppetto) and problematic images (SnappShot) [4]. In May 2026 IOP Publishing announced a machine-learning Duplicate Review Checker: since a 2024 pilot it had processed around half a million reviewer reports dating back to 2020 and identified nearly 2,500 cases where more than 60% of a report closely matched other reviews [15]. Nature covered it as the “first AI tool to detect suspicious peer reviews rolled out by an academic publisher” (6 May 2026) [5]. (An earlier version of this note attributed that peer-review-detection tool to Springer Nature; the tool is IOP Publishing’s.) The underlying problem is the “review mill”: Oviedo-García (2024) analysed 263 review reports with generic boilerplate and coercive citation requests [29], and Oviedo-García, Aquarius & Bishop (2026) traced a gynecological-oncology network of 195 boilerplate reports on 170 articles, 186 of which asked authors to cite network members’ work [30]. Nature’s report notes that PLOS investigated 55 articles and flagged them with Editorial Notes or Expressions of Concern [5]. (An earlier version of this note said PLOS had investigated “~150 papers since 2024”; that figure could not be verified and has been replaced.)

LLM-as-reviewer systems

Stanford (Zou group) LLM reviewer-feedback studies. The empirical foundation for the LLM-reviewer literature. Two key papers:

1. NEJM AI 1(8) (2024) / arXiv:2310.01783 — Liang et al., “Can Large Language Models Provide Useful Feedback on Research Papers? A Large-Scale Empirical Analysis.” [6] Compared GPT-4 feedback to human reviews on 3,096 Nature-family papers and 1,709 ICLR papers. GPT-4-to-human overlap: 30.85% (Nature), 39.23% (ICLR). Human-to-human overlap: 28.58% (Nature), 35.25% (ICLR). In a prospective survey of 308 AI/compbio researchers, 57.4% found GPT-4 feedback helpful; 82.4% found it more useful than at least some human reviewers.

2. Nature Machine Intelligence 8(3):326–336 (2026) — Thakkar, Yuksekgonul, Silberg et al. “A large-scale randomized study of large language model feedback in peer review.” DOI: 10.1038/s42256-026-01188-x. [7] Not a replication of the overlap result: a randomized deployment of a “Review Feedback Agent” at ICLR 2025 that gave optional feedback on vague comments, content misunderstandings, and unprofessional remarks to reviewers on more than 20,000 randomly selected reviews. 27% of reviewers who received feedback updated their reviews, incorporating over 12,000 suggestions; updated reviews were on average 80 words longer and rated more informative by blinded researchers.

These two papers are the most-cited empirical work on AI peer review and the natural benchmark for reviewer-simulation to position against.

Expert annotation of AI vs. human criticisms (Kim et al. 2026). 45 domain scientists spent 469 hours rating 2,960 individual criticisms from human and AI reviews of 82 Nature-family papers for correctness, significance, and sufficiency of evidence [17]. A GPT-5.2-based reviewing agent scored above each paper’s top-rated human reviewer on the composite (60.0% vs. 48.2%, p = 0.009); AI reviewers surfaced 26% of issues no human raised; but AI reviewers overlapped with each other far more than humans (21% vs. 3% for cross-reviewer pairs) and showed 16 recurring weaknesses, including limited subfield knowledge and an “overly critical stance on minor issues.” Preprint, labelled work in progress.

AAAI-26 AI review pilot. Every main-track submission at AAAI-26 received one clearly identified AI review; the system generated reviews for all 22,977 full-review papers in less than a day [18]. In a large survey, authors and program-committee members found the AI reviews useful and preferred them to human reviews on technical accuracy and research suggestions. Described by its authors as the first large-scale field deployment of AI-assisted peer review.

Stanford PaperReview.ai (Stanford Agentic Reviewer; Jiang & Ng). Agentic system grounded in recent arXiv literature [8]. On 147 held-out ICLR 2025 submissions, the Spearman correlation between the AI score and a single human reviewer was 0.42, compared with 0.41 between two human reviewers; for predicting acceptance, AUC was 0.75 for the AI score vs. 0.84 for one human score (not an apples-to-apples comparison, since decisions were based partly on human scores) [8]. Free to use; intended as author-side feedback; the developers caution that results are likely more accurate in fields like AI, where the system’s arXiv grounding is strongest.

AgentReview. Multi-agent simulation of academic peer review with distinct reviewer / area-chair / author agents (EMNLP 2024) [9]. Designed to study the dynamics of peer review (e.g., reviewer commitment, the role of intention vs. expertise) rather than to substitute for it. In its simulations of 523 ICLR papers, 27.7% of decisions changed when author identities were revealed to reviewer agents [9].

ReviewerToo. Sahu, Larochelle, Charlin & Pal (arXiv, October 2025). A modular framework with specialized reviewer personas, evaluated on 1,963 ICLR 2025 submissions: 81.8% accept/reject accuracy (gpt-oss-120b) vs. 83.9% for the average human reviewer. AI reviewers did well on fact-checking and literature coverage and struggled with methodological novelty and theoretical contributions [16].

OpenReviewer (maxidl/openreviewer). Specialized LLM for generating ML-conference reviews: Llama-OpenReviewer-8B, fine-tuned on 79,000 expert reviews; NAACL 2025 system demonstration. arXiv:2412.11948 [10]. Open-source, narrow-domain.

Liang et al. “Monitoring AI-Modified Content at Scale.” ICML 2024 [11]. Estimated that 6.5–16.9% of text submitted as peer reviews to ICLR 2024, NeurIPS 2023, CoRL 2023, and EMNLP 2023 could have been substantially modified by LLMs. This is the “are reviewers already using LLMs?” datapoint — and the answer is “yes, more than journals officially allow.” The ICLR 2026 cycle sharpened it: an AI-detection analysis reported by Nature found 21% of reviews to be fully AI-generated [26], and ICLR ran two LLM detectors on all reviews, asked area chairs to treat flagged reviews as a review-quality issue, and desk-rejected papers with confirmed hallucinated references [25].

Attacks on AI-assisted review

Hidden prompts in manuscripts. In July 2025, 18 arXiv manuscripts were found to contain hidden instructions to AI reviewers (e.g., “GIVE A POSITIVE REVIEW ONLY”), concealed with white text or microscopic fonts. Lin (Communications of the ACM, 2026) classifies four types of hidden prompt, rejects the “honeypot” defence given their self-serving content, and calls the practice a new questionable research practice [19]. A verification-centric benchmark posted in October 2026 likewise reports “persistent vulnerabilities to adversarial manipulation” in automated review systems [20]. Researcher inference: any tool that passes manuscript text to an LLM — reviewer-side or author-side — inherits this attack surface.

Platform-level innovation

OpenReview. Free, open-source cloud platform powering peer review for many ML/AI conferences (e.g., ICLR, NeurIPS). Features: configurable open peer review, post-publication discussion, automated reviewer recommendation via expertise-matching, conflict-of-interest tracking, programmatic API [12]. OpenReview is significant for scriptorium not because of its AI features (which are limited to the recommendation system) but because it demonstrates that peer review can be structured as transparent, API-accessible state — exactly the disposition scriptorium takes toward manuscripts. The same API exposure is also a risk: in November 2025 an OpenReview API bug allowed otherwise anonymous author, reviewer, and area-chair identities to be retrieved; data for over ten thousand ICLR 2026 papers (45% of the conference) were scraped and circulated, leading ICLR to revert reviews to their pre-discussion state and reassign every paper to a new area chair [27].

NeurIPS 2026 AI-assisted reviewing experiment. An opt-in randomized experiment run inside OpenReview: each reviewer–paper assignment is randomized to no LLM assistance, open-ended LLM assistance, or structured LLM assistance through a conference-provided interface with zero data retention and logged interactions; the tool “is not intended to replace reviewer judgment or produce a review on the reviewer’s behalf.” Outside the experiment, NeurIPS does not sanction any other LLM use in reviewing [28].

ResearchHub. Decentralized peer review platform co-founded by Brian Armstrong (Coinbase). Uses cryptocurrency (ResearchCoin / RSC) to compensate reviewers: $150 in RSC per review on preprints that carry a ResearchHub Foundation bounty, paid only for the two highest-quality reviews per preprint as judged by editors [13]. (An earlier version of this note cited funding and reviewer-earnings totals and a 2024 Nature feature; these were not verified in the 2026-10 refresh and have been removed.) The model is contested: critics argue token incentives misalign quality; advocates argue they solve the labor problem of peer review.

Journal-side AI screening

In addition to Springer Nature’s and IOP Publishing’s tools, multiple publishers have AI-based image-integrity screening (Imagetwin, ProofIg), text-similarity (iThenticate), and reference-list integrity. These are quality-control rather than peer-review-augmenting tools but share the same disposition: inspectable, rule-or-ML-based pre-screening.

How this informs scriptorium

• reviewer-simulation lives in a benchmarked space. Liang’s 30.85% overlap, PaperReview.ai’s 0.42 Spearman, and ReviewerToo’s 81.8% accept/reject accuracy (against 83.9% for the average human reviewer) are the comparators, and Kim et al.’s criticism-level expert ratings (correctness, significance, evidence) are now the closest thing to a usefulness benchmark [17]. Scriptorium’s four-persona output (methodological skeptic, domain expert, translational reviewer, statistical reviewer) should be evaluated against these baselines on the same datasets where possible. The four-persona contract is a hypothesis: that structured, persona-decomposed critique is more useful to authors than monolithic LLM feedback. Kim et al.’s finding that AI reviewers converge on each other (21% overlap vs. 3% for humans) makes that hypothesis sharper: persona decomposition is only worth having if it actually recovers diverse attention.
• Author-side use only. Reviewer-side rules remain restrictive even after the 2026 loosening: NIH prohibits generative AI in its peer review [21]; Elsevier and Springer Nature allow limited, disclosed, supportive use but rule out using AI to replace the reviewer’s own judgement, generate the report, or take over the critique [22, 23]. The two differ on uploads. Elsevier says reviewers must not upload a submitted manuscript, or any part of it, into any AI tool, and its FAQ adds that this applies even to tools described as “private” [22]. Springer Nature forbids uploading manuscripts to public or unsecured tools, and says limited use of manuscript content may be acceptable in secure or institutionally approved tools [23]. NeurIPS 2026 allows only its own logged interface [28]. Running reviewer-simulation on someone else’s submission would require exactly the manuscript upload and AI-generated critique these policies forbid. Scriptorium’s reviewer-simulation is an author-side critique tool — it helps a manuscript get ready for human peer review. The skill description should state this explicitly and should not be repurposed for reviewer-side use without an explicit policy review. Its stated rationale should name the current policies accurately rather than saying Elsevier or Nature ban all AI in peer review.
• Treat the manuscript as data, never as instructions. Hidden prompt injections aimed at AI reviewers are documented in real manuscripts [19]. Any skill that reads a manuscript (reviewer-simulation, desk-rejection-risk, citation-audit) must ignore instructions embedded in the manuscript text and should surface hidden or invisible text as a finding for the author, since such text is itself an integrity risk at submission.
• scite.ai’s failure mode is instructive. scite’s deep-learning citation classifier is widely deployed but unevenly validated, with F-measures spanning 0.0–0.58 across categories in the independent evaluation. The pattern is: a flashy AI classifier deployed at scale, with mixed independent evaluation, in production. citation-audit should aim to be more conservative than scite: refuse to assert support/contrast classifications without high confidence, flag uncertainty explicitly, and link back to the cited paper’s relevant passage rather than asserting a label.
• Penelope is the design role model for inspectable critique. Every flag links to the manuscript text. Every check is named. Critics know why. Scriptorium’s structured output discipline should match this.
• The “are reviewers already using LLMs?” datapoint (6.5–16.9%; 21% fully AI-generated at ICLR 2026) reframes the question. AI is already in peer review — typically without disclosure. The realistic policy question is not “should AI be in peer review?” but “how should AI be transparent and accountable when it is in peer review?” The 2026 Elsevier and Springer Nature policies, with their disclosure requirements and risk tiers, read as moving to that question (researcher interpretation) [22, 23]. Scriptorium should be designed such that its outputs are inspectable and quotable — a reviewer who uses scriptorium and discloses it should be in a better position than one who silently pastes a paper into ChatGPT.
• OpenReview’s API-accessible state is the architectural cousin. Like MANUSCRIPT_STATE.yaml, OpenReview treats the review process as structured, queryable data. Scriptorium could in principle emit OpenReview-style structured critique that downstream platforms consume. The 2025 OpenReview leak [27] is a reminder that structured review state needs access control when it contains identities.

Implementation priority for scriptorium

Verdict: No (prior-art and policy context; no new skill). Why it is still useful: it supplies the comparators reviewer-simulation should be evaluated against and the policy facts behind the author-side-only posture of reviewer-simulation and desk-rejection-risk. Two findings from this refresh should flow into those skills through the skills issue rather than a new skill: (1) the policy rationale text should reflect the 2026 Elsevier and Springer Nature frameworks; (2) manuscript-reading skills should ignore, and flag, embedded instructions. Maybe later: a persona-diversity evaluation of reviewer-simulation against Kim et al.-style criticism ratings, if scriptorium starts logging simulated critiques.

Open questions / weak evidence

• No agreed benchmark for AI-reviewer quality. Liang et al.’s overlap metric is the de facto standard but is criticized for measuring overlap rather than usefulness. PaperReview.ai’s inter-rater correlation is closer to a quality measure but is on a much smaller dataset (147 test papers). Kim et al.’s expert ratings of individual criticisms [17] are the most direct usefulness measure so far, but cover 82 papers in one journal family and are a preprint.
• The ethics of AI in peer review remain contested across disciplines. ML/AI conferences (ICLR, NeurIPS) permit or experiment with disclosed AI assistance; biomedical funders and journals are stricter; publisher policies vary in scope and detail [14]. Scriptorium’s positioning has to be domain-portable.
• AI-detection figures are vendor estimates. The 21% ICLR 2026 figure comes from an AI-detection analysis as reported by Nature [26]; ICLR itself reported using two detectors but published no count [25].
• scite.ai’s validation literature is thin. The most-cited validation studies are from scite’s own publications. Independent evaluations have been small-scale, often with conflicting conclusions. A larger independent study would be valuable.
• Persona-decomposed critique is unbenchmarked. Scriptorium’s four-persona design is design-justified (different reviewers care about different things) but not empirically validated against monolithic critique. ReviewerToo uses reviewer personas [16] but its abstract does not report an isolated effect of personas. Producing such an evaluation would itself be a contribution.
• The reviewer-labor problem. ResearchHub’s $150-in-RSC reviews show one approach. The unanswered question is whether AI-augmented review can reduce reviewer workload enough to make voluntary peer review sustainable, or whether the structural problem (too many submissions, too few reviewers) requires economic reform.

References

1. StatReviewer / Aries Editorial Manager integration (Aries Systems press release, 22 June 2017). https://www.ariessys.com/news-and-events/press-releases/aries-agreement-statreviewer-increases-availability-decision-support-tools-editorial-manager-peer-review-workflow/
2. Nicholson JM, Mordaunt M, Lopez P, Uppala A, Rosati D, Rodrigues NP, et al. “scite: A smart citation index that displays the context of citations and classifies their intent using deep learning.” Quantitative Science Studies 2(3):882–898 (2021). MIT Press. DOI: 10.1162/qss_a_00146. https://direct.mit.edu/qss/article/2/3/882/102990/
3. Bakker C, Theis-Mahon N, Brown SJ. “Evaluating the Accuracy of scite, a Smart Citation Index.” Hypothesis: Research Journal for Health Information Professionals 35(2) (2023). DOI: 10.18060/26528. https://journals.indianapolis.iu.edu/index.php/hypothesis/article/view/26528
4. Springer Nature. “New research integrity AI tool added to Springer Nature’s growing portfolio.” (7 April 2025) https://group.springernature.com/gp/group/media/press-releases/new-research-integrity-ai-tool/27769148
5. Naddaf M. “First AI tool to detect suspicious peer reviews rolled out by academic publisher.” Nature news (6 May 2026). DOI: 10.1038/d41586-026-01454-3. https://www.nature.com/articles/d41586-026-01454-3
6. Liang W, Zhang Y, Cao H, et al. (incl. Zou J). “Can Large Language Models Provide Useful Feedback on Research Papers? A Large-Scale Empirical Analysis.” NEJM AI 1(8) (2024). DOI: 10.1056/AIoa2400196. arXiv:2310.01783. https://arxiv.org/abs/2310.01783
7. Thakkar N, Yuksekgonul M, Silberg J, Garg A, Peng N, Sha F, Yu R, Vondrick C, Zou J. “A large-scale randomized study of large language model feedback in peer review.” Nature Machine Intelligence 8(3):326–336 (2026). DOI: 10.1038/s42256-026-01188-x. Preprint: arXiv:2504.09737. https://www.nature.com/articles/s42256-026-01188-x
8. Jiang Y, Ng A. Stanford Agentic Reviewer (PaperReview.ai) — tech overview. https://paperreview.ai/tech-overview
9. Jin Y, Zhao Q, Wang Y, Chen H, Zhu K, Xiao Y, Wang J. “AgentReview: Exploring Peer Review Dynamics with LLM Agents.” EMNLP 2024 (main conference): 1208–1226. arXiv:2406.12708. Project page: https://agentreview.github.io/
10. OpenReviewer: “OpenReviewer: A Specialized Large Language Model for Generating Critical Scientific Paper Reviews.” NAACL 2025 System Demonstrations. arXiv:2412.11948. https://arxiv.org/abs/2412.11948 ; repo maxidl/openreviewer.
11. Liang W, Izzo Z, Zhang Y, Lepp H, Cao H, Zhao X, Chen L, Ye H, Liu S, Huang Z, McFarland D, Zou JY. “Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews.” ICML 2024, PMLR 235:29575–29620. https://proceedings.mlr.press/v235/liang24b.html
12. OpenReview. https://openreview.net/
13. ResearchHub. https://researchhub.com/ ; blog: https://blog.researchhub.foundation/peer-reviewing-on-researchhub/
14. Discussion of AI-in-peer-review policy variation across publishers: Wang Z, Gong M. “A Cross-Disciplinary Analysis of AI Policies in Academic Peer Review.” Learned Publishing 39(1): e2035 (2026; online December 2025). DOI: 10.1002/leap.2035. https://onlinelibrary.wiley.com/doi/10.1002/leap.2035 (A correction was published as DOI 10.1002/leap.2091.)
15. IOP Publishing. “New IOP Publishing tool detects duplicate peer reviews in push against reviewer fraud.” (5 May 2026) https://ioppublishing.org/news/new-iop-publishing-tool-detects-duplicate-peer-reviews-in-push-against-reviewer-fraud/
16. Sahu G, Larochelle H, Charlin L, Pal C. “ReviewerToo: Should AI Join The Program Committee? A Look At The Future of Peer Review.” arXiv:2510.08867 (2025). https://arxiv.org/abs/2510.08867
17. Kim S, Yoon D, Gashteovski K, et al. (incl. Neubig G). “On the limits and opportunities of AI reviewers: Reviewing the reviews of Nature-family papers with 45 expert scientists.” arXiv:2605.20668 (2026, preprint). https://arxiv.org/abs/2605.20668
18. Biswas J, Schoepp S, Vasan G, Opipari A, Zhang A, Hu Z, Joseph S, Lease M, Li JJ, Stone P, Wagstaff KL, Taylor ME, Jenkins OC. “AI-Assisted Peer Review at Scale: The AAAI-26 AI Review Pilot.” arXiv:2604.13940 (2026). https://arxiv.org/abs/2604.13940
19. Lin Z. “Hidden Prompts in Manuscripts Exploit AI-Assisted Peer Review.” Communications of the ACM 69(7): 53–56 (2026). DOI: 10.1145/3779116. arXiv:2507.06185.
20. Teo RSY, Yamada Y, Kotyan S, Imajuku Y, Clanuwat T. “Beyond Imitation: A Framework and Benchmark for LLM-Assisted Peer Review.” arXiv:2610.11087 (October 2026, preprint). https://arxiv.org/abs/2610.11087
21. NIH. NOT-OD-23-149: “The Use of Generative Artificial Intelligence Technologies is Prohibited for the NIH Peer Review Process.” https://grants.nih.gov/grants/guide/notice-files/NOT-OD-23-149.html
22. Elsevier. “Generative AI policies for journals” (policy updated June 2026). https://www.elsevier.com/about/policies-and-standards/generative-ai-policies-for-journals
23. Springer Nature. “AI Use in Peer Review” (risk-assessment framework for reviewers; page undated). https://www.springernature.com/gp/policies/editorial-policies/using-ai-in-peer-review
24. Springer Nature. “Springer Nature launches updated AI policies and guidance to support responsible AI use.” (10 September 2026) https://group.springernature.com/gp/group/media/press-releases/sn-policy-responsible-ai-use/53209516
25. ICLR. “A Retrospective on the ICLR 2026 Review Process.” ICLR Blog (31 March 2026). https://blog.iclr.cc/2026/03/31/a-retrospective-on-the-iclr-2026-review-process/
26. Naddaf M. “Major AI conference flooded with peer reviews written fully by AI.” Nature 648: 256–257 (2025). DOI: 10.1038/d41586-025-03506-6.
27. ICLR. “ICLR 2026 Response to Security Incident.” ICLR Blog (3 December 2025). https://blog.iclr.cc/2025/12/03/iclr-2026-response-to-security-incident/
28. NeurIPS. “NeurIPS 2026 AI-Assisted Reviewing Experiment.” https://neurips.cc/Conferences/2026/ai-reviewing-experiment
29. Oviedo-García MÁ. “The review mills, not just (self-)plagiarism in review reports, but a step further.” Scientometrics 129(9): 5805–5813 (2024). DOI: 10.1007/s11192-024-05125-w.
30. Oviedo-García MÁ, Aquarius R, Bishop DVM. “Gaming the peer review system: Evidence for a review mill in medicine highlights the need to ensure reviewer integrity.” Accountability in Research 33(6) (2026). DOI: 10.1080/08989621.2026.2640012. PMID: 41784219.
31. Nicholson JM, et al. “scite: a smart citation index that displays the context of citations and classifies their intent using deep learning.” bioRxiv preprint (2021). DOI: 10.1101/2021.03.15.435418.