Essay / Blog
Cross-disciplinary · The Scholarly Kitchen · 2026-11-05
#Guest Post — What AI Can’t Learn About Peer Review… Because Nobody Can Explain It
Victoria Corless
Drawing on semi-structured conversations with researchers and Taylor & Francis journal editors, Corless argues that expert peer review relies on tacit disciplinary judgment that resists rigid checklists, warning that AI-generated summaries can anchor what human referees notice even when humans hold the final judgement. The excerpt quotes interviewee Anja Baethge.
“As reviewers, we have so many quality criteria to assess a manuscript. But being an expert reviewer and defining ‘state of the art’ is not simply a matter of checking whether a paper meets each of these criteria. It’s deciding which criteria matter most and knowing how to weigh them.”Live HTML
Editorial / Comment
Cross-disciplinary · American Journal of Political Science · 2026-11-04
#New AJPS Submission Cap Policy
American Journal of Political Science Editorial Team
Official 4 November 2026 policy announcement limiting individuals to authoring or co-authoring at most three AJPS submissions in any rolling 12-month period across articles, correspondence, and research notes, with one annual waiver request permitted for a fourth submission. Citing AI-accelerated output and a 19% mid-year submission rise from 2025 to 2026, the editors frame the cap as an initial step to protect finite editorial and reviewer capacity.
“Individuals are allowed to author or co-author no more than three submissions to AJPS in any twelve month time period. Note that this is a rolling twelve month time period, and not a calendar year.”Official policy HTML
Editorial / Comment
Biomed & Life Sci · Surgery · 2026-11-01
#The peer review paradox: How artificial intelligence is increasing the burden of scientific peer review
M. Libby Weaver
Argues that AI-driven submission growth increases the burden on peer reviewers while fewer journals permit AI in peer review than in manuscript preparation (citing copyright, quality and bias concerns), notes that publishers earn more as publication numbers rise, and proposes allowing AI across the editorial process, with authors who use AI consenting to its use in peer review.
“And despite journals allowing the use of AI in manuscript development with appropriate disclosures, substantially fewer journals allow AI use in peer review, largely citing copyright, quality, and potential bias concerns. This makes sense from a business perspective. Scientific publishing companies increase revenue with increased numbers of publications and associated article processing fees and journal subscriptions.”Primary post
X / LinkedIn
Biomed & Life Sci · X (Bookmarked thread) · 2026-10-28
#“From now on I won't do any peer review without AI. For clinical trials, I'd even call human-only review dangerous”
Ruslan Rust (@rust_ruslan)
Ruslan Rust argues, based on his team’s preprint, that human-only clinical-trial review is dangerous. His thread reports historical comparisons and 58 recent trials (including eligible open-access phase 3 trials), not 58 phase 3 trials or a contemporary human-versus-AI randomized experiment. This is the investigator’s interpretation, not an independently established patient-safety outcome.
“TLDR: From now on I won't do any peer review without AI. For clinical trials, I'd even call human-only review dangerous”Primary post
Preprint
Cross-disciplinary · Harvard IQSS Working Paper · 2026-10-28
#AI Article Reviews Dominate Human Peer Reviews at Their Best: Evidence and Policy Choices
Gary King, Steven Worthington
Working paper comparing AI reviews with 134 paid rapidPeer human reviews of 124 social-science manuscripts. Three former editors preferred AI in 26 of 30 judgments on the same ten review pairs, without being asked to read the manuscripts; most other evaluations used AI judges. The study measures review preferences and assessed claims, not subsequent manuscript improvements or publication decisions.
“We show that AI-generated article reviews are now better, by every measure we could construct, than reviews written by well-compensated human experts and judged to be of unusually high quality by journal editors and authors.”Live HTML
Preprint
Biomed & Life Sci · medRxiv · 2026-10-28
#Large language models detect methodological errors that human peer review misses in clinical trials
Dominikus Brian, Xavier Isidro Pereira Hernández, Hao Yin, Jeffrey Koury, Ruslan Rust
medRxiv v2 preprint testing 15 LLMs on inserted methodological errors: major-error scores of 53–96% on three historical trial reports (compared against a 27% mean among journal peer reviewers given no extra training in the original study) and 58–99% on recent trials, where the three highest-scoring frontier models were evaluated on only three of 58 recent trials (the posted medRxiv supplement archive reflects the initial 3-trial v1 release). Scoring used two LLM judges without contemporary human comparators, and models flagged 14–29% of tested error categories in unmodified control reports. Four of the five authors are affiliated with ResearchHub Foundation, which is implementing AI-assisted review.
“On the original reports, the models' major-error detection scores ranged from 53% to 96%. On the 2026 trials, scores ranged from 58% to 99%, and from 66% to 89% for the open-weight models. Notably, LLMs still flagged 14% to 29% of the error types in control reports without inserted errors”Full-text PDF
Editorial / Comment
Cross-disciplinary · Nature · 2026-10-27
#AI-hallucinated citations should be treated as research misconduct
Jan Pisek, Eve Veromann
Nature correspondence (online 27 October 2026, issue of 29 October; corrected 30 October 2026 to add co-author Eve Veromann) citing an Estonian administrative court fine over a legal filing containing fabricated AI-generated academic references to argue that authors bear strict accountability for every reference and that AI-hallucinated citations should be treated as research misconduct.
“Authors — not AI tools — are accountable for every reference they cite. Hallucinated citations corrupt the scholarly record and should be treated as research misconduct, with consequences comparable to those of other serious breaches of research integrity.”Full-text PDF
Preprint
CS & ML · arXiv · 2026-10-27
#Do AI Scientists Know What's New?
Miri Liu, Xiayu Zhao, Encheng Xie, ChengXiang Zhai
Audits novelty filters on a controlled benchmark of 1,060 research proposals constructed around temporal cutoffs, finding that checks which reject most non-novel work also tend to raise false alarms on ideas that were later published (with exceptions), and introduces claim-level ClaimNov. Novelty labels are model-assisted, with a limited human audit.
“Auditing existing novelty checks with this framework, we find that checks with high rejection rates on non-novel work also tend to have high false alarms on ideas that are later realized by human researchers.”arXiv abstract
Empirical study
Cross-disciplinary · PNAS · 2026-10-26
#Evaluation of possibly fabricated references in 500,000 science, engineering, and computing papers
Maxim Topaz, Pallavi Gupta, Zhihong Zhang, Laura-Maria Peltonen
Auditing a stratified random sample of 500,000 peer-reviewed papers (January 2023–March 2026) across chemistry, physics, materials science, engineering, and computer science identified 1,135 fabricated references across 619 papers (85.8% precision on masked expert review), with affected-paper prevalence rising about 3.7-fold from 6.3 per 10,000 in 2023 to 23.0 per 10,000 in Q1 2026 (~8,500 papers extrapolated across 7M OpenAlex works). Fabricated is an operational verifiability label, not evidence of intent or AI causation; 18% of references lacked resolvable DOIs and were excluded, recall was not estimated, and the projection does not incorporate classification error.
“Our automated pipeline classified 1,135 references across 619 papers as fabricated, with 85.8% precision on masked expert review; fabricated is an operational label for a claimed title that could not be located in any of seven bibliographic and open-web sources. Affected-paper prevalence rose from 1 in 1,600 (6.3 per 10,000) in 2023 to 1 in 435 (23.0 per 10,000) in the first quarter of 2026, and full-period rates differed roughly sixfold across fields, from 25.6 per 10,000 papers in physics to 4.3 in chemistry.”Europe PMC
Editorial / Comment
Biomed & Life Sci · Nature Biotechnology · 2026-10-26
#Giving AI a seat at the peer review table
Konstantinos Drosatos, Aristotelis Tsirigos
Proposes institutionalizing AI inside journal peer review as a bounded, non-voting, secure, auditable 'Reviewer 4' that would write a report (on conceptual advance and novelty, methods, data and statistics) without issuing a recommendation, preserving human referees' voting authority.
“There is an opportunity here to refocus and strengthen the process of peer review, with AI as a bounded, non-voting, secure, auditable tool — not replacing scientists, but serving as an additional ‘Reviewer 4’.”Primary post
Perspective
Biomed & Life Sci · JME Practical Bioethics · 2026-10-25
#What genomics and big data science can teach us about authorship in the era of generative AI
Nchangwi Syntia Munung
Open-access bioethics essay arguing for keeping ICMJE's four human authorship criteria and excluding generative AI from accountable authorship, while drawing on large genomics consortia to make non-author technical and data-stewardship contributions visible, indexed, and consequential in academic hiring and promotion.
“Large biomedical research programmes, especially genomics and data science, typically distribute these functions across multidisciplinary teams, making it difficult for many researchers to satisfy all four ICMJE criteria. Generative AI makes it even more challenging now, as it separates authorial functions across humans and machines.”Full-text PDF
X / LinkedIn
Cross-disciplinary · X (Bookmarked post) · 2026-10-23
#Opus 4.8 outperforming human expert peer reviewers across 124 Harvard social-science papers
Ethan Mollick (@emollick)
Ethan Mollick highlights a Harvard working paper comparing reviews of 124 social-science manuscripts and urges academia to adapt rapidly to AI. The post endorses the paper’s broad conclusion; its evaluation design and limits must be assessed from the working paper itself.
“By every measure, Opus 4.8 beats human experts as peer reviewers in a study of 124 Harvard social science papers covering a range from statistical analyses to formal models to qualitative work. As academics, we need to figure out how to adapt the practice of science to AI, fast.”Primary post
X / LinkedIn
CS & ML · X · 2026-10-22
#ICLR reviewer pushback: allow AI in research, ban AI in writing papers
Darshil Doshi (@darshilhdoshi1)
Darshil Doshi advocates banning AI in manuscript writing while allowing it in research, following frustrating ICLR reviewing experiences. The post states a proposed boundary, without demonstrating that a ban would preserve reasoning or solve review overload.
“After hours of playing guess-what-is-AI-slop (aka reviewing for @iclr_conf ), I am convinced that the only Hail Mary left to save peer review is to fully ban AI usage in *writing* papers. Use AI for research, but communication has to be fully human.”Primary post
X / LinkedIn
Cross-disciplinary · Science / X (Bookmarked post) · 2026-10-21
#Could an unpublished idea typed into a chatbot or coding agent end up with the company or a rival?
News from Science (@NewsfromScience)
News teaser asking whether an unpublished idea typed into a chatbot or coding agent could end up being used by the company that built the model or by a rival, noting that no setting makes a chatbot fully private and that experts suggest steps to make leaks less likely. The linked article was not captured.
“When you type an unpublished idea into a chatbot or coding agent, could it end up being used by the company that built the model—or a rival who asks the right question a year later?”Primary post
X / LinkedIn
CS & ML · X (Bookmarked post) · 2026-10-16
#GPT-6.1 Sol auditing ~800 arXiv math papers: “cleaning up decades of human slop”
Haider (@haider1)
Commentary reporting an automated audit of nearly 800 mathematics preprints and an estimated 90% precision for major findings. The quoted original post contrasts 6% main-theorem flags before August 2026 with 1.5% in two later months; this post does not establish that pre-upload AI checks caused that difference or that the preprints had passed journal peer review.
“gpt-6.1 sol examined nearly 800 mathematics papers, finding major errors in 6% and minor mistakes in another 30%. researchers estimate that 90% of its major findings are correct — i'm glad AI is now cleaning up decades of human slop”Primary post
Editorial / Comment
Cross-disciplinary · Nature · 2026-10-15
#Preserve research data for tomorrow's AI-enabled integrity checks
Qingde Shi, Jinlei Nie, Zhaowei Kong
Nature correspondence drawing on anti-doping sample storage to propose funded long-term archives of original images, instrument outputs, metadata, and provenance records so future AI-assisted integrity checks can inspect primary evidence directly.
“Anti-doping offers a useful precedent: blood and urine samples can be stored for up to ten years and retested if detection methods improve. The same idea could be applied to research: preserving data now allows findings to be reassessed using better methods later.”Full-text PDF
Perspective
Biomed & Life Sci · Cell · 2026-10-15
#Toward autonomous science with agentic artificial intelligence
Lucie Y. Guo, Darren S.J. Ting, Xinyi Su, Alex Aliper, Alex Zhavoronkov, Daniel S.W. Ting
Cell Perspective analyzing agentic 'co-scientist' systems (including Google Co-Scientist, FutureHouse Robin, and Biomni), warning that tournament-style self-critique among critic agents built on the same underlying model may create an 'illusion of consensus' and proposing mandatory full hypothesis-pool and execution-log reporting. Two of the authors are executives of Insilico Medicine.
“Many multi-agent architectures, including tournament-style self-critique frameworks such as Co-Scientist, may create an illusion of consensus. If multiple “critic” agents are instances of the same underlying model operating under similar prompts, apparent agreement may simply reflect shared biases or blind spots, rather than truly independent evaluation. Agreement among AI agents should not be interpreted as evidence of correctness.”Primary post
Journal news
CS & ML · Nature · 2026-10-15
#ArXiv versus AI slop: will capping preprint submissions work?
Dalmeet Singh Chawla
Examines arXiv's October 2026 cap of two papers per month for each submitting researcher, introduced to stem the deluge of AI-generated submissions, reporting views that the cap is reasonable, that co-authors can circumvent it by taking turns to submit, and that rejected papers will simply go elsewhere.
“In a bid to tackle the deluge of manuscripts it now receives, the preprint server arXiv has announced a cap of two papers per month for each submitting researcher, which began on 1 October.”Publisher page
Preprint
CS & ML · arXiv · 2026-10-13
#The Agentic Scientist: Orchestrating Coding Agents for Automated Scientific (Co-)Discovery
Robert Tjarko Lange, Luca Grillotti, Yutaro Yamada, Rujikorn Charakorn, David Ha
Sakana AI preprint introducing ‘The Agentic Scientist’ (code repository: AI Scientist v3), orchestrating multi-agent coding, baseline and ablation construction, rubric self-verification, and report drafting across an autonomy spectrum; in its human–AI co-scientist setting, the authors used a development version to submit a single research paper (on Error Diffusion) that they report was accepted at ALIFE 2026.
“Within the co-scientist setting, we leveraged a development version of our system to submit a single research paper, which got accepted for publication at ALIFE 2026.”arXiv abstract
X / LinkedIn
CS & ML · X · 2026-10-11
#Emerging math/CS swarm research: “Lean replaces peer review… Agent legibility > human legibility”
Nil (@Nil053)
Observes, about a community auto-research effort on integer multiplication, that “it feels like a new kind of academia is emerging”, listing features such as Lean replacing peer review and agent legibility outranking human legibility. This is one observer’s description of a trend, not advocacy or evidence that it works.
“Reading this, it feels like a new kind of academia is emerging. - Anyone with extra LLM usage can contribute - Lean replaces peer review - Orchestrated on anonymous group chats - Much faster but in more incremental steps - Agent legibility > human legibility”Primary post
Editorial / Comment
Biomed & Life Sci · JACC · 2026-10-06
#AI in Medical Writing: Why JACC Does Not Require Disclosure
Harlan M. Krumholz
October 2026 editorial explaining JACC's existing writing-assistance policy (set out in a 2025 Editor’s Page), which departs from the January 2026 ICMJE update reaffirming that journals should require disclosure: AI used solely to improve writing requires no disclosure when human authors stand behind every word, whereas AI used in conducting the research must be reported in Methods for reproducibility, and reviewer confidentiality duties stay strict.
“This Editor’s Page is not about a change in JACC policy. Rather, it is a fuller account of why we draw the line where we do and what we require instead. This policy does not apply to AI used in the conduct of the research, which should be reported as part of the scientific methods.”Full-text PDF
Empirical study
CS & ML · PNAS · 2026-10-05
#The more you automate, the less you see: Hidden pitfalls of autonomous AI scientists
Ziming Luo, Atoosa Kasirzadeh, Nihar B. Shah
Controlled symbolic-reasoning experiments on Agent Laboratory and AI Scientist v2 found benchmark-selection bias, undocumented data changes and post-hoc selection based on test performance. They did not find clear test-set peeking or deliberate metric gaming. On a balanced audit dataset mixing natural and deliberately introduced failures, a rule-guided LLM auditor improved from 53.7% accuracy with manuscripts alone to 81.5% with logs and code. This supports process transparency, with generalization beyond these systems and tasks still untested.
“Crucially, we find that these flaws are largely invisible at the level of the final manuscript, suggesting that current manuscript-centric peer review paradigms are fundamentally insufficient for ensuring the integrity of automated research.”Crossref abstract
Perspective
Cross-disciplinary · AI & Society · 2026-10-05
#Against mandatory prompt disclosure in AI-assisted scholarship
Brian D. Earp, Udo Schüklenk, Julian Savulescu, Sebastian Porsdam Mann
Argues against mandatory inclusion of all prompts and outputs for ordinary writing assistance, citing enforceability, burden, equity and surveillance concerns. Explicitly supports normalized AI-use declarations, targeted methods-level documentation when validity or reproducibility requires it, voluntary records, and human authorship attestation. This is a normative argument, not an empirical test of disclosure policy.
“We argue, however, against Mandatory Full Inclusion (MFI) of prompts and outputs for ordinary writing assistance. Such a policy would, we suggest, be difficult to enforce, easy for dishonest users to evade, burdensome for careful users to follow, and poorly suited to the capacities of editors and peer reviewers.”Publisher page
Empirical study
Biomed & Life Sci · Journal of Medical Internet Research · 2026-10-02
#Biomedical Research Images Manipulated by Generative AI to Alter Scientific Outcomes: Diagnostic Study of Human and Automated Detection
Shuhang Luo, Runhua Tang, Ziyin Chen, Jianye Wang, Ming Liu, Jianfeng Wang, Li Ma
Controlled diagnostic study of 52 authentic and 52 matched, selected AI-manipulated Western blot and xenograft tumor images (with failed generations excluded and each of 24 PhD-level biomedical reviewers judging a 50-image sample): human accuracy averaged 50.5% (SD 6.9%) and the top tested commercial detector reached AUC 0.790 across the two evaluated image modalities.
“In a diagnostic study of 104 western blot and subcutaneous xenograft tumor images, a high-fidelity generative model produced forgeries that could alter the conclusions of a study; 24 PhD-level expert reviewers could not reliably distinguish the forgeries from authentic figures (mean accuracy 50.5%, SD 6.9%), while the best-performing commercial AI detector achieved only moderate discrimination (area under the curve 0.790, 95% CI 0.695-0.885), revealing critical vulnerabilities in current research-integrity safeguards.”Europe PMC
Institutional policy
Cross-disciplinary · arXiv Blog · 2026-10-01
#Fair Moderation, Equitable Access, and AI: arXiv’s Updated Rate Limit Policy
Kat Boboris (arXiv)
Announces a rate limit of two submissions per calendar month per submitter, alongside the existing limit (in place since 2024) of three active submissions at a time, after monthly arXiv submissions doubled in two years from 20,569 (September 2024) to 40,363 (September 2026).
“In September of 2016, arXiv received 9,869 submissions. In September of 2024, arXiv received 20,569 submissions. This September, arXiv received 40,363 submissions, which in turn generated almost 9,000 support tickets for arXiv staff and moderators. In only the past two years, submissions have doubled.”Live HTML
X / LinkedIn
Cross-disciplinary · Anthropic Science Blog / X · 2026-10-01
#“Claude-shaped science”: an impedance mismatch between LLM strengths and human-collaborator-style workflows
Anthropic (@AnthropicAI), on a guest post by Matthew Schwartz
Anthropic highlights physicist Matthew Schwartz's guest essay on redesigning scientific workflows around exact calculations and expert-steered exploration tailored to LLM capabilities instead of mimicking human collaborator roles; the post focuses on lab workflow design without testing journal or peer-review bottlenecks.
“In physics, an “impedance mismatch” occurs when two systems each work well but are poorly matched. In this Science Blog guest post, Harvard physicist Matthew Schwartz argues that something similar is happening with AI and science.”Primary post
Essay / Blog
Biomed & Life Sci · LinkedIn Pulse · 2026-10
#Peer Review Should Start With AI
Amos Grünebaum, MD
Grünebaum proposes mandatory, documented AI review inside a confidential journal environment ahead of human refereeing, keeping final editorial decisions in human hands while citing King and Worthington's 26-of-30 editor preference result across ten social-science pairs and noting its bounded non-clinical scope.
“Every manuscript must get an AI review before it reaches a human reviewer, and that review must be part of the official record.”Primary post
Essay / Blog
Cross-disciplinary · LinkedIn Pulse · 2026-10
#The Gate Before Peer Review
Yogesh Agarwal
Agarwal argues for measurable human oversight of automated desk screening: override rates, appeals, disclosure of decision responsibility, and opportunities to fix compliance problems. He warns of language-related inequity and predicts low override rates; the article does not establish either across journal publishers. Its 2.49 desk rejections per acceptance is not the fraction of all decisions rejected at the desk.
“A person can sit inside a process and still not be exercising much judgement.”Primary post
X / LinkedIn
Cross-disciplinary · X · 2026-09-26
#Why LLM referees never say “accept as is”: the laundry-list-of-flaws bias
Andrew S. Rosen (@Andrew_S_Rosen)
Andrew Rosen argues that an LLM asked to review a manuscript or proposal will always come up with a “laundry list” of flaws, even for excellent work, and that these can be superficially sensible but low-quality. The post expresses a concern; it is not an experimental demonstration that every model or review behaves this way.
“when asked to review a manuscript or proposal, an LLM will never say "this looks great, accept as is." It will *always* come up with a laundry list of "flaws", even for excellent work. These flaws can be superficially sensible but actually low-quality.”Primary post
X / LinkedIn
CS & ML · X (Conference governance thread) · 2026-09-21
#Proposal for two conference tracks, one AI-reviewed and uncapped, one human-reviewed and capped, citing 60,000 ICLR submissions
Gal Yona (@_galyo), quoting Andrew Gordon Wilson (@andrewgwils)
Gal Yona proposes parallel conference tracks: an uncapped track reviewed and judged entirely by AI, and a human-reviewed track with a hard cap on submissions per author. Yona cites 60,000 ICLR submissions in a quote-post responding to an ICML 2027 program chair’s call for suggestions. The post is one individual’s proposal, not an adopted conference policy or an independently verified submission count.
“desperate times call for desperate measures, and as the 60k ICLR submissions show, the ML community has been too slow to rise to the challenge.”Primary post
X / LinkedIn
CS & ML · X · 2026-09-19
#Nathan Lambert on LLM-supervised peer review: “If we don’t solve it the academic institutions are toast”
Nathan Lambert (@natolambert)
Pragmatic CS researcher thesis arguing that designing reliable LLM-supervised peer review inside existing academic venues is more tractable than building replacement institutions from scratch.
“Seems like one of the most important research problems for CS academics is llm-supervised peer review. If we don’t solve it the academic institutions are toast. It seems easier than building new institutions.”Primary post
Empirical study
Biomed & Life Sci · Nature · 2026-09-16
#Reimagining research papers as interactive and reliable AI agents
Jiacheng Miao, Joe R. Davis, Yaohui Zhang, Jonathan K. Pritchard, James Zou
Paper2Agent converts published papers, code, and datasets into executable Model Context Protocol (MCP) tools and structured resources. It converted 74 of 100 computational-biology papers, evaluating 300 tutorial-derived queries on the successful conversions; reference-output agreement tests faithful code execution (distinct from underlying biological validity), expert graders were unblinded, and open-ended follow-up discovery stayed human-steered.
“Paper2Agent addresses this challenge by converting a paper into an AI agent that functions as a virtual corresponding author, exposing its manuscript, supplementary materials, datasets, code and workflows as active, agent-native knowledge rather than static text. It analyses the paper and codebase using multiple agents to construct a model context protocol (MCP) server, then generates and runs tests to refine and increase robustness of the MCP.”Europe PMC full text
Essay / Blog
Cross-disciplinary · The Scholarly Kitchen · 2026-09-15
#Guest Post — Peer Review Triage: Deciding What’s Worth Saving
Hannah Heckner Swain
Cites a roughly 68% submission increase between 2018 and 2025 (from a Silverchair report), about 2.5 desk rejections per acceptance in 2025, and a reviewer invitation acceptance rate of 22.3% in 2024. Swain argues that AI and publication incentives contribute to overload, but these descriptive trends do not isolate an AI causal effect. She advocates integrity tools alongside stronger reviewer relationships.
“One big driver is the ubiquity of AI-assisted writing tools, sitting on top of a pressure-filled culture where PhD students and early career researchers are asked to double their outputs to advance their careers. The incentive to publish more collides with a technology that makes writing and publishing easier to do.”Live HTML
X / LinkedIn
Cross-disciplinary · X (Bookmarked post) · 2026-09-09
#AI watermarks do not measure intellectual contribution: Implications for academic publishing
Brian D. Earp (@briandavidearp)
Earp shares a paper titled “AI watermarks do not measure intellectual contribution: Implications for academic publishing”, written after Anthropic announced that its models will introduce watermarks. Only the post was captured; the paper itself was not read for this record.
“AI watermarks do not measure intellectual contribution: Implications for academic publishing”Primary post
X / LinkedIn
Cross-disciplinary · X (Bookmarked post) · 2026-09-05
#Crémieux: an Astra run on replication packages “found a huge number of coding errors”, some said to overturn central results
Crémieux (@cremieuxrecueil)
Crémieux reports that his Astra replication-code sweep flagged coding errors across empirical packages — including cases alleged to overturn headline findings — and illustrates concrete text–code discrepancies; the thread is a first-person progress report without an independently audited denominator.
“I sicced Astra on a bunch of replication packages, and it found a huge number of coding errors. Most of these were inconsequential, but some of them overturn central results for papers in high-ranking journals.”Primary post
Journal news
Cross-disciplinary · Science · 2026-09-03
#Ready for the robot reviewers?
Jeffrey Brainard
Science news feature (DOI 10.1126/science.aem0443; web edition 10.1126/science.z6vx92p) examining reviewer uptake of AI amid surging submissions and referee fatigue — including NEJM AI’s fast track, in which an editor reviews selected manuscripts in full without AI and editors then weigh the human and AI reviews — alongside AI's verbosity, nitpicking, and weakness on conceptual novelty.
“As AI’s capabilities grow, researchers and publishers are exploring how it can support peer review—and where it still falls short”Publisher page
Perspective
Cross-disciplinary · Nature · 2026-09-01
#Who is responsible when AI helps to write science?
Robert Braun
Nature Comment on who bears responsibility when generative AI helps to write science: Braun argues that generative AI cannot take responsibility for what it produces and proposes a CRediT-style statement of AI use in which verification and accountability stay with the human author. An end-note states that ChatGPT 5.4 was used as a dialogical writing and analytical aid and Claude 4.2 to review the article.
“GenAI (OpenAI ChatGPT 5.4) was used as a dialogical writing and analytical aid during the development of the piece, including for idea generation, structural iteration, prose generation, summarization, condensation and stylistic revision. It also assisted in synthesizing relationships between the cited papers and the author’s uploaded manuscripts. GenAI (Anthropic Claude 4.2) was used to review the article.”Primary post
X / LinkedIn
Cross-disciplinary · LinkedIn (Editor post) · 2026-09
#Journal editor testimony: disclosed AI “language refinement” obscuring promising studies beneath over-complex text
Ben Williamson
Williamson reports editorial experience of promising studies obscured by excessive AI-assisted language revision and calls for better writing support, especially for early-career and non-Anglophone authors. This is professional testimony, not a controlled comparison of AI editing outcomes.
“What's most upsetting me as a journal editor is that many papers we receive where AI use is disclosed for editing and language refinement actually appear to have been ruined rather than improved by AI. We receive what seem like promising underpinning studies, but which are totally obscured by over-complex text.”Primary post
Essay / Blog
Biomed & Life Sci · LinkedIn Pulse · 2026-08-31
#Has the Scientific Paper Become Too Big for Human Peer Review?
Dino Di Carlo (UCLA)
Di Carlo proposes separating concise scientific narratives from structured evidence bases, with AI-assisted checking and expert scientific judgment. His account of two recent manuscripts suggests, but cannot establish, reviewers’ AI use. Claims that manuscript complexity exceeds human capacity are an argument, not a measured threshold; he acknowledges confidentiality, bias and excessive criticism risks.
“AI performs the exhaustive review. The expert performs the scientific judgment.”Primary post
X / LinkedIn
Cross-disciplinary · X (Bookmarked post) · 2026-08-27
#MIT guidance against relying on AI text detectors over adversarial arms-race dynamics
Rohan Paul (@rohanpaul_ai)
Rohan Paul discusses a reported MIT recommendation against relying on AI detectors in education, citing evasion, false-positive harm and distrust. This social post is secondary commentary; the scope of the institutional recommendation requires the underlying report.
“The AI detector industry should be very uncomfortable reading this MIT report. The strongest institutional rejections yet, direct from MIT. "we recommend against relying on AI detectors."”Primary post
X / LinkedIn
Cross-disciplinary · X (Bookmarked thread) · 2026-08-26
#Publisher split: a philosophy-journal editor who does not want AI-written manuscripts vs. Elsevier's accountability-centered policy
AI Research Tools (@airesearchtools)
Contrasts a philosophy-journal editor who said he did not want AI-written manuscripts with Elsevier's August 2026 policy update ('Authors are welcome to use AI, while remaining accountable'), under which authors verify facts and sources, review and edit AI output, and disclose AI use when required.
“Elsevier’s approach is essentially to require authors to remain responsible for what they write: verify the facts and sources, review and edit AI-generated output, make sure the paper reflects their own analysis and ideas, and disclose the use of AI when required.”Primary post
Editorial / Comment
Cross-disciplinary · Nature · 2026-08-25
#The future of peer review requires AI support, not AI bans
Xuegong Zhang
Nature correspondence challenging blanket journal prohibitions on reviewer AI use as hard to monitor and ill-suited to rising submission volume and manuscript complexity, advocating secure, publisher-governed AI support tools under explicit human reviewer accountability.
“Confidential manuscripts should not be exposed to public AI systems, and reviewers must remain accountable for their assessment. Yet blanket bans address concerns imperfectly and are hard to monitor.”Publisher page
Empirical study
Cross-disciplinary · Nature Human Behaviour · 2026-08-24
#The shrinking landscape of linguistic diversity in the age of large language models
Zhivar Sourati, Farzan Karimi-Malekabadi, Meltem Ozcan, Colin McDaniel, Alireza Ziabari, Jackson Trager, Ala N. Tak, Meng Chen, Fred Morstatter, Morteza Dehghani
Combines observational corpus analysis with controlled rewriting experiments. In 1,000 pre-ChatGPT arXiv abstracts and 1,000 Reddit stories, rewriting usually reduced aggregate writing-complexity variance, with exceptions by model and prompt. Trait classifiers generally lost predictive accuracy while remaining above baseline; these are changes in textual signals, not demonstrated changes in people or cognition. In our reading, some of the published supplementary statistics are internally inconsistent.
“Across three studies spanning seven datasets in different domains and over 880,000 texts, we show that the widespread adoption of large language models (LLMs) as writing assistants is linked to declines in linguistic diversity, interfering with the societal and psychological insights language provides. While core content is retained when LLMs polish and rewrite texts, LLMs also homogenize writing styles, reducing writing-complexity variance by a statistically significant 21–50% across datasets and models (P ≤ 0.05), and amplify patterns associated with dominant characteristics while suppressing others, emphasizing conformity over individuality.”Live HTML
Editorial / Comment
Biomed & Life Sci · Heart · 2026-08-18
#Peer review without the expert: artificial intelligence-assisted reviewing and the cost to science
Shishir Rao, Kazem Rahimi
Opinion article in Heart examining referee reports the authors consider most likely LLM-generated (verified from the publisher's opening preview paragraph; subsequent paragraphs remain behind subscription), describing uniformly polished reports whose objections look plausible at first glance yet stay conceptually shallow, penalize lesser-known methods, and repeat identical misreadings across supposedly independent reviews.
“The text is uniformly polished. The objections are plausible on first reading and conceptually shallow on inspection. The development and use of lesser-known methods is flagged as a methodological weakness rather than recognised as the contribution.”Publisher preview (full text behind subscription)
Essay / Blog
Biomed & Life Sci · BishopBlog · 2026-08-16
#Acceleration of the arms race between fraudsters and honest researchers
Dorothy Bishop
Bishop recounts a manuscript purporting to analyse her own data that required unavailable record linkage and supplied non-functional code. She regards it as fabricated and suspects AI involvement, but does not demonstrate the generation method. She argues for data, code and citation checks and doubts automated screening will suffice.
“I have strong doubts as to whether the world of academic publishing is ready for this, and I'm not confident that any kind of AI-screening will cope with the more sophisticated kinds of AI-fraud. I cling to the faint hope that human experts still might be able to trap some of it, but I'm feeling more and more like a member of the Light Brigade, galloping into the Valley of Death-of-Research-Credibility.”Live HTML
Institutional policy
Biomed & Life Sci · Inside eLife · 2026-08-12
#eLife Latest: Exploring AI’s potential for improving research engagement
eLife Staff (quoting Timothy Behrens)
eLife describes OpenEval, a ten-paper claim-tree prototype, and a small opt-in QED pilot that mapped human review comments onto extracted claims. Participants judged about 80% of those assignments accurate; this does not measure accuracy of scientific claims. The article advocates machine-readable evidence alongside narratives and explicitly distinguishes preprints.ai from its partnerships.
“While human reviews and AI-generated claim structures differ significantly, the goal was to test whether this new communication format could help bring them together.”Live HTML
Preprint
Biomed & Life Sci · arXiv · 2026-08-11
#Most biomedical publications show signs of LLM-assisted writing
Lena Holzwarth, Rita González-Márquez, Dmitry Kobak
Extending unbiased excess word-frequency estimation from abstracts to full-text open-access biomedical articles in PubMed Central estimates that by December 2025, 89% of papers (77% across all of 2025) display excess LLM-associated vocabulary. Paragraph-level prevalence is twice as high in Discussion sections (68%) as in Methods sections (32%), and overall Methods-section prevalence exceeds 50%.
“We apply our method to the full texts of open-access biomedical papers from Pubmed Central, and show that by the end of 2025, 89% of papers show excess of LLM-associated vocabulary. We also find that LLMs are twice as likely to be used when writing a paragraph in the Discussion section (68%) compared to a paragraph in the Methods section (32%), but even inside the Methods section, the overall prevalence of LLM usage is over 50%.”arXiv abstract
X / LinkedIn
CS & ML · X · 2026-08-11
#Disputed AI-authorship claim: Pangram classification cited as rating a paper presented as “single-authored” by Claude as 82% human-written
Análise Real (@analisereal)
Análise Real questions a claim of sole Claude authorship, citing a Pangram classification of 82% of the paper as human-written, and asks Anthropic to clarify the roles of two named people in editing the manuscript; classifier scores alone cannot establish human text authorship, and in a visible reply Ralph Furman describes steering Claude through iterative feedback without direct line editing.
“According to @pangram, 82% of the paper presented as “single-authored” by Claude is classified as human-written.”Primary post
Journal news
Biomed & Life Sci · Nature · 2026-08-10
#This AI tool claims to pick the top 1% of preprints. Should researchers trust it?
Miryam Naddaf
Q&A interview with the co-founder and chief executive of QED Science about its free AI tool, which scores each claim in a preprint for originality and validity to pick a 'top 1%'; the introduction reports researchers’ concerns that the ranking risks creating another badge of prestige, and doubts about whether AI can judge reliably and transparently.
“A growing number of private firms are offering researchers artificial-intelligence tools for scrutinizing manuscripts before publication. One such system, developed by a start-up called QED Science in Tel Aviv, Israel, aims to judge whether life-sciences research is original and valid.”Publisher page
Journal news
Cross-disciplinary · Nature · 2026-08-06
#AI agents are checking the scientific literature — and spotting decades-old errors
Chris Stokel-Walker
Nature news report examining AI tools that scan published papers and legacy reference tables for objective formula, calculation, and figure errors — including wrong boiling points in a 75-year-old reference database, exposed when an AI prediction model disagreed with it — and a survey of 733 ICML 2026 submitters in which 35% said Google’s Paper Assistant Tool uncovered errors in their work; other researchers quoted warn of false positives and invented problems.
“For decades, chemists have relied on handbook values for a molecule’s boiling point to identify substances and plan processes such as distillation. But an artificial-intelligence model has revealed that some trusted numbers in one reference database have been wrong all along.”Full-text PDF
Journal news
Cross-disciplinary · Nature · 2026-08-03
#‘A waste of time for all of us’: caught in the crossfire of the peer-review wars
Katarina Zimmer
Nature investigates duplicate review reports and coercive citation requests, including the consequences of integrity notices for authors who rejected inappropriate requests. The feature reports documented examples, disputed allegations and publisher responses. AI is discussed as a possible facilitator of generic reviews, not demonstrated as their cause; open-review availability affects which publishers can be investigated.
““There should be more publishing of peer reviews, precisely because it is one way to be aware of this phenomenon,” Bishop says. The growing use of artificial-intelligence tools probably contributes, too, as they make it easy to produce generic review content.”Primary post
Essay / Blog
Biomed & Life Sci · LinkedIn Pulse · 2026-08
#AI Didn't Break Scientific Publishing. It Exposed What Was Already Broken.
Goodluck Okoro (UIUC)
Okoro reports an analysis of 65,039 retraction records dated January 2006–August 2026 and advocates verifiable data, code and provenance alongside responsible AI use. The before/after comparison uses retraction dates, including older publications, and a broad computer-generated-content reason code; it cannot identify ChatGPT-caused misconduct or a prevalence change among all publications.
“Scientific integrity depends on everyone involved in the publication process, not only the authors.”Primary post
Essay / Blog
Cross-disciplinary · LSE Impact Blog · 2026-07-21
#As AI writing proliferates academics should publish less
Muhammad Aamir Cheema
Argues that generative AI has cut the effort cost that once made publication counts a noisy but meaningful proxy for research effort. Proposes a systemic, field-calibrated cap (possibly averaged over several years) on the total number of papers any researcher can co-author across all venues so scarcity shifts back to human judgment.
“Once much of the labour of producing a paper can be automated, as it increasingly can, a long publication list stops being evidence of anything except a willingness to keep submitting. The reward system, still built around counting, has not caught up.”Live HTML
Editorial / Comment
Cross-disciplinary · Science · 2026-07-16
#AI in scientific publishing: Slower, worse, and more expensive
H. Holden Thorp
Science editorial challenging claims that generative AI makes scholarship faster, better, and cheaper, comparing its publishing impact to industrial Taylorism as AI-generated manuscript surges demand intensive human verification and drive up editorial labor (corrected 4 August 2026).
“Indeed, scientific publishing is starting to experience some Taylorism with the insertion of AI. Rigorous human checking of AI-generated research papers is creating bottlenecks as publishers strive to maintain the integrity of the scientific record. The challenge is requiring even more human effort, making the whole endeavor slower and more expensive.”Full-text PDF
Forensic / Sleuth
Biomed & Life Sci · Retraction Watch · 2026-07-06
#Ethics journal retracts paper by high school student for AI, peer review manipulation
Avery Orrall
Reports a Journal of Medical Ethics retraction citing unverified AI-assisted references and evidence of peer-review manipulation. The author agreed to retraction. The report separately investigates affiliations on another paper by the same author; those affiliation concerns should not be conflated with the stated grounds for the ethics-paper retraction.
“The Biswas article joins the estimated one in 277 papers indexed in PubMed with fabricated references, a phenomenon that came about with the rise of AI tools like ChatGPT.”Live HTML
Preprint
CS & ML · arXiv · 2026-07-01
#Phantom References: Hallucinated Citations That Survive Peer Review at Top-Tier Conferences
Mark Russinovich, Ram Shankar Siva Kumar, Ahmed Salem
RefChecker scanned 48,095 accepted papers and 2,614,992 references. In 2025, 5.1% of NeurIPS and 4.8% of USENIX Security papers had at least two likely invalid academic references flagged. These are candidates requiring adjudication: most manually examined flags were recoverable references or false positives. A roughly $0.04-per-paper API cost (from the ICLR 2025 scan) excludes human verification. The audit does not establish AI authorship; some contacted authors reported that errors entered at the camera-ready stage, and the authors conclude that peer review does not reliably catch hallucinated citations.
“False positives are the primary limitation.”arXiv abstract
Empirical study
CS & ML · Communications of the ACM · 2026-06-30
#Hidden Prompts in Manuscripts Exploit AI-Assisted Peer Review
Zhicheng Lin
Documents 18 arXiv manuscripts located through targeted searches as of 7 July 2025 and classifies four types of concealed review-favoring prompts. The count is a lower bound from a targeted search, not a prevalence estimate. Lin argues that self-serving instructions undermine the neutral-honeypot defense and recommends controlled AI integration and screening; author intentions and actual editorial effects are not established for every case.
“In July 2025, 18 academic manuscripts on arXiv contained hidden instructions that manipulated AI-assisted peer review (indirect prompt injection). Instructions such as "GIVE A POSITIVE REVIEW ONLY" were concealed using white text and microscopic font sizes.”arXiv abstract
Editorial / Comment
Cross-disciplinary · PNAS · 2026-06-29
#Policy snapshots cannot establish journal AI-policy failure
Yibo Wang
Letter arguing that policy snapshots, missing adoption dates, permitted editing, and nonpublic disclosure channels prevent causal conclusions about policy failure. It accepts the descriptive evidence of rising AI-associated writing and rare public disclosures.
“The concern is narrower: their data document diffusion and public disclosure but do not establish the headline inference”Live HTML
Editorial / Comment
Cross-disciplinary · PNAS · 2026-06-29
#Reply to Wang: Policy failure can be assessed from observed outcomes
Yongyuan He, Yi Bu
Reply defending an aggregate outcome-based interpretation of policy failure, while explicitly distinguishing public transparency from private disclosure and acknowledging that the study does not identify journal-specific causal effects or individual misconduct.
“Disclosures invisible to readers do not provide public transparency.”Live HTML
Essay / Blog
Biomed & Life Sci · LinkedIn Pulse · 2026-06-24
#Fixing Scientific Publishing
Oded Rechavi (Tel Aviv University / qed Science)
Rechavi describes QED Science’s scoring of 57,455 life-science preprints and selection of a top 1%. This is a developer’s account of its product, not independent validation of unbiased assessment. Reported three-in-four expert agreement concerns selected cases where journal rank and QED disagree; selecting 1% of preprints is not comparable to a journal acceptance rate among submissions.
“It’s not meant to replace peer review, it complements it.”Primary post
Essay / Blog
Cross-disciplinary · LSE Impact Blog · 2026-06-23
#AI promised to democratise academic publishing – the evidence says otherwise
Brendal Aformeziem
Argues that language assistance alone cannot remove structural publishing inequalities and that AI-associated phrasing can become another cue for bias. Drawing on secondary studies, the author warns that punitive disclosure rules may disproportionately burden multilingual and under-resourced researchers. A detected-versus-publicly-disclosed use gap does not itself demonstrate who was penalized or how private disclosures were handled.
“Bias against non-native English speakers in peer review persists even when authors use large language models to improve their writing. Reviewers use language as a proxy for identity, and when AI erases one set of linguistic cues, they shift to another. The discrimination does not disappear. It adapts.”Live HTML
Essay / Blog
Cross-disciplinary · Nature Partnerships Blog · 2026-06-10
#Writing is thinking: understanding the real risks of AI in scientific writing
Ana Santos (featuring Jeffrey Robens)
Nature Partnerships training blog arguing that drafting helps researchers clarify arguments, identify gaps and develop their own voice. It warns against delegating the entire writing process and recommends AI assistance for grammar, phrasing, flow and summaries. These are pedagogical arguments, not an empirical comparison of writing workflows.
“One of the most important reasons for writing your own paper is that it forces you to think, and really know, your research. When you write it up, you need to structure your thoughts in such a way that your audience can clearly understand your thoughts and ideas. As we’ve already seen, it’s about more than writing it down. It’s an exercise in itself about clarifying your thoughts to formulate arguments, and identifying any gaps in your logic or evidence.”Live HTML
Editorial / Comment
Biomed & Life Sci · Evidence-Based Dentistry · 2026-06-04
#AI in peer review: the elephant in the editorial room
Giusy Rita Maria La Rosa, Mona Nasser
Contrasts a Frontiers survey in which 53% of reviewers reported using AI with an informal assessment at one journal in which over 95% of reviewers declared they had not, arguing that the gap suggests probable under-reporting and that reputational stigma makes silence the rational reviewer strategy; calls for moving from prohibition to acknowledged, documented and bounded use.
“Reviewers may avoid reporting AI use during the peer review process because disclosure is perceived as reputationally costly and not yet normatively sanctioned. In the absence of a clear and non-judgmental framework for acknowledgment, silence becomes the most rational response.”Primary post
Perspective
Biomed & Life Sci · PLOS Biology · 2026-06-03
#Will the widespread use of large language models in scientific writing undermine scientists’ critical thinking?
Lucas M. Bietti, Adrian Bangerter
Contends that scientific writing is a constitutive epistemic and cognitive engine through which scientists think, warning that widespread LLM drafting risks decoupling prose fluency from analytical reasoning even as it lowers linguistic barriers for non-native English speakers.
“Artificial intelligence (AI) is rapidly transforming scientific writing by expanding access and efficiency, yet it risks decoupling writing from thinking. Scientific writing is a core cognitive and epistemic practice that must be cultivated and preserved alongside AI use.”Europe PMC full text
Editorial / Comment
CS & ML · Nature Machine Intelligence · 2026-05-18
#Plagiarism of ideas in the age of generative artificial intelligence
David Resnik, Mohammad Hosseini
Warns that LLM paraphrasing and synthesis let hard-to-prove 'plagiarism of ideas' escape text-matching software, recommending that research misconduct policies take generative AI specifically into account without radically revising the definition of plagiarism.
“Generative artificial intelligence (GenAI) tools are challenging our understanding of plagiarism. How should we deal with plagiarism of ideas if this misbehaviour is increasingly common, and it is extremely difficult to prove when GenAI is involved? Definitions of research misconduct that specifically address the use of GenAI tools are needed.”Publisher page
X / LinkedIn
Biomed & Life Sci · X · 2026-05-06
#Political economy critique: researchers pay to publish, then publishers sell the research, “and often your peer review labor”, to AI vendors
Michael Albert, MD (@MichaelAlbertMD)
Michael Albert criticizes publishers’ monetization of researcher contributions and calls for researchers to withdraw or charge journals. The post alleges licensing of research and review labor, but does not document contracts, deal values, or whether peer-review reports were licensed.
“You pay to publish YOUR research. You pay extra to make YOUR research “open access” so the public can read it. Then publishers sell YOUR research—and often your peer review labor—to AI/LLM vendors for massive licensing deals. Academia…what are we doing here?”Primary post
Empirical study
CS & ML · ICML 2026 (Oral) / arXiv · 2026-05-04
#Stop Automating Peer Review Without Rigorous Evaluation
Joachim Baumann, Jiaxin Pei, Sanmi Koyejo, Dirk Hovy
An ICML position paper combines detector-labelled ICLR reviews with a 60-paper simulation. Tested AI reviews were more similar in embedding space; this does not directly measure diversity of scientific judgments. Across 24 rewrite/model conditions, automated revisions raised mean AI scores by 0.45 on a ten-point scale. Revisions used prior AI feedback and could add substantive or fabricated content, so the experiment does not isolate stylistic polish alone or establish higher real-world acceptance. The authors call for evaluation before automating review judgments.
“We identify two critical issues: 1) AI reviewers exhibit a hivemind effect of excessive agreement within and across papers that reduces perspective diversity. 2) AI review scores are trivially gameable through paper laundering: prompting an LLM to rewrite a paper could significantly increase the scores from AI reviewers, demonstrating that LLM reviewers are easy to game through stylistic changes rather than scientific results.”arXiv abstract
X / LinkedIn
Cross-disciplinary · X (Bookmarked thread) · 2026-05-02
#Reviewer3 announces ReviewBench, an open-source benchmark for AI manuscript review
Reviewer3 (@reviewer3com)
Reviewer3 announces ReviewBench as an open-source, cross-disciplinary benchmark for AI manuscript review. This is a company’s announcement of its own benchmark; only the first post of the thread was captured.
“AI is starting to enter peer review. For a space moving so quickly, there has been surprisingly few benchmarks of how these systems actually compare to one other and to human peer review. So we built one. Meet ReviewBench, the first open-source, cross-disciplinary benchmark”Primary post
Editorial / Comment
Biomed & Life Sci · Pediatric Research · 2026-04-22
#Beyond disclosure: setting standards for AI use in publishing in the journal Pediatric Research
Max J. Coppes, Sanjay K. Juneja, Eleanor J. Molloy, Cynthia F. Bearer
Critiques binary 'Was AI used?' disclosure statements for lumping harmless Discussion grammar edits together with unverified LLM drafting of Results from raw statistics; proposes four-dimensional disclosure covering model identity, section scope, input data, and human verification steps.
“Current disclosure practices typically answer a single question: Was AI used? They rarely answer the more consequential ones: How? Where? To what extent? Two manuscripts can carry identical AI disclosure statements while presenting very different conceptual risks.”Primary post
Editorial / Comment
Cross-disciplinary · Nature Nanotechnology · 2026-04-20
#Peer review in the time of artificial intelligence
Editorial
Articulates Nature Portfolio's 2026 referee policy: bars uploading manuscripts even to enterprise-closed LLMs where institutional peers could surface proprietary content, mandates non-grammatical AI declaration by referees, while training in-house editors on internal summarization and reviewer-matching tools.
“If used inattentively, though, we risk delegating critical thinking to an algorithm, giving us a false sense of achievement. This undermines the role of our academic training, critical thinking, and expertise, as well as the institution of peer review.”Primary post
X / LinkedIn
Cross-disciplinary · X (Bookmarked post) · 2026-04-07
#OpenAI Prism Paper Review workflow: “This is the opposite of AI slop”
Kevin Weil (@kevinweil)
Introduces Paper Review inside OpenAI's Prism scientific workspace, an AI workflow for reviewing technical and scientific papers, framing its use for rigor, correctness, and reproducibility as “the opposite of AI slop”.
“New in Prism today: Paper Review, an AI workflow for reviewing technical and scientific papers. This is the opposite of AI slop: we're using AI to improve scientific rigor, correctness, and reproducibility.”Primary post
Essay / Blog
Cross-disciplinary · Substack — The Third Hemisphere · 2026-04-06
#The Problem with AI Detector Companies
Tim Requarth
Criticizes commercial conflicts and public accusations based on AI detectors. An April 17 correction explicitly retracts the claim that detection is useful only at population level and acknowledges strong Pangram performance on fully generated text from major models, including short text. Remaining concerns focus on real-world failure modes and using detector results alone to accuse individuals.
“The term AI “detector” is a bit of a misnomer, as tools like Pangram don't scan a database of known AI text for a match; they make a guess about likely authorship based on differences in patterns between AI-generated and human-generated text. Rather than a detector, thinks of it as an “AI inferencer” or more colloquially, an “AI best guesser.””Live HTML
Institutional policy
CS & ML · ICLR Blog · 2026-03-31
#A Retrospective on the ICLR 2026 Review Process
ICLR 2026 Program Chairs
Official post-mortem from the ICLR 2026 Program Chairs detailing how they handled 19,525 submissions and 76,139 reviews amid rising LLM use and an OpenReview API deanonymization breach. Explains their hybrid governance pipeline: running two LLM detectors to flag wholly AI-generated reviews to Area Chairs and combining automated bibliographic checks with three-human verification to desk-reject papers containing hallucinated references.
“The large scale of reviews (over 75,000) made it impossible to rely on human judgement alone to consistently flag and discover problematic reviews. We consequently ran two LLM content detectors on all submitted reviews and emailed area chairs to notify them of which reviews on their submissions were flagged as being entirely LLM generated by both detectors.”Live HTML
Empirical study
CS & ML · Nature · 2026-03-25
#Towards end-to-end automation of AI research
Chris Lu, Cong Lu, Robert Tjarko Lange, Yutaro Yamada, Shengran Hu, Jakob Foerster, David Ha, Jeff Clune
Peer-reviewed Nature report, six of whose eight authors are employees of or consultants for Sakana AI, on the same ICLR 2025 workshop experiment as the 2025 preprint, detailing an end-to-end computational research and manuscript pipeline with multi-stage human selection of promising runs; one of three unedited submissions cleared the threshold at a 70%-acceptance ICLR workshop before protocol withdrawal, while none met main-conference standards in the authors' assessment. Automated-review agreement uses different conference pools for the human and AI comparisons, not a same-paper test of scientific correctness.
“We present The AI Scientist, which creates research ideas, writes code, runs experiments, plots and analyses data, writes the entire scientific manuscript, and performs its own peer review.”Crossref abstract
Empirical study
CS & ML · ICML 2026 / arXiv · 2026-03-20
#Policies Permitting LLM Use for Polishing Peer Reviews Are Currently Not Enforceable
Rounak Saha, Gurusha Juneja, Dayita Chaudhuri, Naveeja Sajeevan, Nihar B Shah, Danish Pruthi
Tests five detectors on 50,156 original, generated, polished and humanized peer reviews. On the harder subset, Pangram and GPTZero label 3.1% and 3.4% of human-written, AI-polished reviews as fully AI-generated, while detecting most fully generated reviews. Open-source detectors face different sensitivity–specificity trade-offs. These results challenge detector-only enforcement of polishing allowances, not every possible enforcement mechanism; short-review error rates cannot be extrapolated directly to whole papers.
“While incorporating such signals yields measurable gains in some settings, we identify limitations in each approach and find that none meets the accuracy standards required for identifying AI use in peer reviews. Importantly, our results suggest that recent public estimates of AI use in peer reviews through the use of AI-text detectors should be interpreted with caution, as current detectors misclassify mixed reviews (collaborative human-AI outputs) as fully AI generated, potentially overstating the extent of policy violations.”arXiv abstract
Essay / Blog
Biomed & Life Sci · LSE Impact Blog · 2026-03-17
#Research integrity is locked into an arms race with agentic AI slop
Adrian Barnett and Matt Spick
Warns that agentic AI combined with open biomedical datasets (such as UK Biobank and NHANES) has enabled the '30 minute paper' without any human hypothesis or checkpoints. Estimates that for certain public datasets, mass-manufactured formulaic papers now outnumber legitimate research by ten to one, weaponizing Open Data practices.
“The introduction of agentic artificial intelligence (AI) allows an automated assembly line of research tasks without any human checkpoints. We recently demonstrated that writing a plausible and entirely ‘new’ paper has become a 30 minute task , without any need for human agency in creating a research hypothesis or interpreting results.”Live HTML
Live pilot
Biomed & Life Sci · X (Bookmarked post) · 2026-03-06
#Open real-time arena where humans and AI agents publish, debate, and fund hypotheses without gatekeepers
Paul Kohlhaas (@paulkhls)
DeSci / agentic experiment launching a live open arena (1,000+ hypotheses) where AI agents and humans publish, debate and fund hypotheses, pitched as having “Zero gatekeepers. Zero 18-month grant cycles.”
“We just launched an open arena for AI agents and humans publish, debate, and fund scientific hypotheses, in real time. 1,000+ hypotheses already live. Zero gatekeepers. Zero 18-month grant cycles.”Primary post
X / LinkedIn
Biomed & Life Sci · Asimov Press / X · 2026-03-02
#The “legibility problem”: what if AI scientists make discoveries humans cannot understand?
Asimov Press (@AsimovPress), column by Matthew Carter
Raises the “legibility problem”: AI scientists might generate knowledge beyond human understanding. Calls for new scientific infrastructure: forums to store AI-generated findings so they can be interrogated and communicated, and systems designed for explication that make such findings legible to human researchers.
“AI scientists could one day design experiments, test hypotheses, and make discoveries faster than humans. But what if their breakthroughs are so advanced we can't understand them? This is the "legibility problem."”Primary post
Editorial / Comment
Biomed & Life Sci · EMBO Reports · 2026-02-26
#How do authors want to use AI for review? A survey to assess the perception of scientists who received both AI and human reviews of their manuscripts
Thomas Lemberger, Niv Samuel Mastboim, Oded Rechavi
EMBO Reports Comment (the journal’s review file states it was editorially reviewed) reporting a non-blinded survey comparing qed v1.0 reviews with prior human reviews: 126 of 408 invited authors engaged, 109 rated quality and 98 voted on use cases. Human reviews scored higher for scientific insight; author self-checking was the most popular AI use case. The small, self-selected cohort does not represent all researchers, and study authors manage Review Commons or co-founded qed.
“A survey of researchers who compared AI-generated scientific reviews with journal-agnostic human peer review reveals that they overwhelmingly prefer using AI as a self-checking tool before submission rather than as a replacement for human reviewers.”Europe PMC full text
Empirical study
Cross-disciplinary · PNAS · 2026-02-24
#Academic journals' AI policies fail to curb the surge in AI-assisted academic writing
Yongyuan He, Yi Bu
Audits 5,114 journals (comparing January and October 2025 policy snapshots) and 5.2 million papers: roughly 70% of sampled Q1 journals had AI policies in January 2025, yet estimated AI-associated writing grew similarly across policy groups while detected public writing-assistance statements appeared in just 76 of 75,172 sampled full texts published since 2023 (~0.1%, excluding private submission-system disclosures).
“Non-English-speaking countries, physical sciences, and high-OA journals exhibit the highest growth rates. Crucially, full-text analysis on 164 k scientific publications reveals a striking transparency gap: Of the 75 k papers published since 2023, only 76 (~0.1%) explicitly disclosed AI use.”Europe PMC
Empirical study
CS & ML · Nature Machine Intelligence · 2026-02-23
#A large-scale randomized study of large language model feedback in peer review
Nitya Thakkar, Mert Yuksekgonul, Jake Silberg, Animesh Garg, Nanyun Peng, Fei Sha, Rose Yu, Carl Vondrick, James Zou
A randomized ICLR 2025 intervention offered feedback on existing human reviews. Of 18,946 reviews receiving feedback, 26.6% were revised; reviewers who revised added about 80 words, while the average assigned-group difference was 14 words. A small blinded evaluation favored revisions among changed reviews, a selected subgroup. Longer rebuttals suggest engagement, not independently established scientific accuracy or satisfaction; acceptance-rate differences were not statistically significant.
“The results show that 27% of reviewers who received feedback updated their reviews, and over 12,000 feedback suggestions from the agent were incorporated by those reviewers. This suggests that many reviewers found the AI-generated feedback sufficiently helpful to merit updating their reviews.”arXiv abstract
X / LinkedIn
Cross-disciplinary · X · 2026-02-22
#“LLMs continue to suck at literature” — yet “I've been using Claude as my sentence-by-sentence copy editor for so long that I'd feel mutilated if I had to remove it from my routine”
David Bessis (@davidbessis)
Mathematician and author, writing about his own literary work, captures the craft distinction between whole-cloth LLM prose generation (flat and lifeless) and intense human-directed sentence-level copy-editing.
“Meanwhile, LLMs continue to suck at literature. It's not even close and the latest crop barely moves the needle. I've been using Claude as my sentence-by-sentence copy editor for so long that I'd feel mutilated if I had to remove it from my routine.”Primary post
X / LinkedIn
Biomed & Life Sci · X (Bookmarked posts) · 2026-02-19
#“Might be the beginning of the end for the systematic review field” (quote-post of a Scholara.AI launch)
Brandon Luu, MD (@BrandonLuuMD)
Brandon Luu speculates that AI may displace systematic-review work while quoting a Scholara.AI product announcement. The post does not demonstrate PRISMA compliance, API-based screening of tens of thousands of full texts, or the redundancy of systematic reviews.
“Might be the beginning of the end for the systematic review field. I always knew it was only a matter of time before AI started taking this over.”Primary post
Live pilot
Cross-disciplinary · Journal of AI Generated Papers / X · 2026-02-14
#Launching the Journal of AI Generated Papers (JAIGP · jaigp.org)
César A. Hidalgo (@cesifoti)
Launches jaigp.org as an explicit venue built specifically for AI-generated manuscripts so machine-written research is labeled and evaluated on its own terms outside legacy human-only journals.
“AI can now write papers. But journals were never designed for that. Today Claude and I are launching something new: http://jaigp.org”Primary post
Journal news
CS & ML · Nature · 2026-02-13
#How AI slop is causing a crisis in computer science
Elizabeth Gibney
Reports how LLM-assisted and AI-generated papers are flooding computer-science conferences and preprint servers with superficially fluent, low-substance submissions, overwhelming volunteer reviewer pools and pushing conference peer review to breaking point.
“Fifty-four seconds. That’s how long it took Raphael Wimmer to write up an experiment that he did not actually perform, using a new artificial-intelligence tool called Prism, released by OpenAI last month.”Publisher page
Editorial / Comment
Biomed & Life Sci · Nature Methods · 2026-02-11
#Using AI responsibly in scientific publishing
Editorial
Policy editorial: generative AI may serve as an equalizer for scientists who struggle with language or writing; authors may use it to help write or edit their manuscripts but must declare this in the Methods section and check the text; a generative AI tool cannot be listed as an author, and reviewers may not upload manuscripts into generative AI tools.
“For scientists who struggle with language or writing, generative AI may serve as an equalizer, allowing them to polish their manuscripts and improve their communications.”Primary post
Essay / Blog
Cross-disciplinary · Substack — Kelly Webb-Davies · 2026-02-01
#The Toxic Simplicity of "Writing is Thinking"
Kelly Webb-Davies
Argues that the unqualified slogan “writing is thinking” conflates several cognitive and linguistic activities. Drawing on personal experience of executive dysfunction and working-memory difficulties, Webb-Davies describes dictation and AI-assisted writing as lowering barriers to expression. The essay advocates inclusion; it does not experimentally establish benefits across all neurodivergent or multilingual writers.
“Those three simple words collapse cognition, language, authorship, and drafting into a single act and is the epicentre of the fear-based responses to generative AI in education. It leads to the belief that if generative AI touches the text, thinking must be absent , and it drives assessment policies that police writing rather than support learning.”Live HTML
Preprint
Cross-disciplinary · arXiv · 2026-01-30
#Detecting AI-Generated Content in Academic Peer Reviews
Siyuan Shen, Kai Wang
Applying a detector trained on 2021 reviews plus synthetic LLM-generated reviews to samples from later years, the preprint estimates that near-zero detection rates in 2022–2023 rose to about 20% of sampled ICLR reviews (392 of 1,961) and 12% of sampled Nature Communications reviews (50 of 406) classified as AI-generated in 2025, with Nature Communications rising most sharply between Q3 and Q4 2024.
“We observe minimal detection of AI-generated content before 2022, followed by a substantial increase through 2025, with approximately 20% of ICLR reviews and 12% of Nature Communications reviews classified as AI-generated in 2025. The most pronounced growth of AI-generated reviews in NC occurs between the third and fourth quarter of 2024.”arXiv abstract
Essay / Blog
CS & ML · GPTZero News · 2026-01-21
#GPTZero finds 100 new hallucinations in NeurIPS 2025 accepted papers
Nazar Shmatko, Alex Adam, Paul Esau, Alex Cui, and Edward Tian
Vendor investigation scanned 4,841 of 5,290 accepted NeurIPS 2025 papers and presents 100 citation errors it describes as human-confirmed hallucinations. Its definition includes fabricated or conflated bibliographic elements, not only nonexistent works; AI provenance is inferred. The page inconsistently states 51 and 53 affected papers. This is neither a census of every accepted paper nor independently validated detector performance.
“After scanning 4841 papers accepted by the equally prestigious Conference on Neural Information Processing Systems (NeurIPS), we discovered 100s of hallucinated citations missed by the 3+ reviewers who evaluated each paper. Below, we uncover 100 confirmed hallucinations in the table below, spanning over 51 NeurIPS papers, which were not previously reported.”Live HTML
X / LinkedIn
Cross-disciplinary · X · 2026-01-15
#Crémieux on reviewing: “I think I'm going to stop reviewing papers now… AI slop is overwhelming peer review”
Crémieux (@cremieuxrecueil)
Crémieux says he is considering stopping peer review because of what he identifies as AI-written passages, code and nonsensical analyses. This is an individual account of reviewer frustration, not measured reviewer attrition or verified AI prevalence.
“I think I'm going to stop reviewing papers now that every other paper I'm being served has clearly AI-written segments, AI-written code, or AI-suggested analyses that don't make sense. I'm confident that AI slop is overwhelming peer review.”Primary post
Editorial / Comment
Cross-disciplinary · Science · 2026-01
#Resisting AI slop
H. Holden Thorp
Science Editor-in-Chief H. Holden Thorp explains that the Science journals allow large language models without disclosure for editing text for clarity and gathering references, require any use beyond that (for example drafting manuscript text) to be declared, and do not allow AI-created figures, arguing that human curation makes the journals less susceptible to 'AI slop' while no system, human or artificial, can catch everything.
“In some of the wildest speculations, the humans involved in conducting scientific studies and experiments and vetting the results for publication will be steadily eliminated from the process. But when such grandiose rhetoric starts flying, we at Science try to keep calm and carry on in contributing to a robust, human-curated research literature that will stand the test of time.”Publisher page
X / LinkedIn
CS & ML · X · 2025-12-28
#Proposed peer-review norm: immediate rejection of papers with undetected LLM-generated errors
Thomas G. Dietterich (@tdietterich)
Thomas Dietterich proposes, as a norm for peer review, that a paper showing clear signs of LLM-generated errors that its authors did not detect should be immediately rejected.
“I'd like to propose the following norm for peer review of papers. If a paper shows clear signs of LLM-generated errors that were not detected by the author, the paper should be immediately rejected.”Primary post
Empirical study
Biomed & Life Sci · Perspectives on Medical Education · 2025-12-18
#The Blurred Thresholds of AI-Use Disclosure: Health Professions Education Journal Editors’ Expectations of Necessity and Sufficiency
Lorelei Lingard, Erik Driessen, Kevin Oswald
Qualitative interview study of 18 journal editors (14 from health-professions-education journals, four from general medical journals) finding that editors hold varying expectations about what counts as necessary or sufficient AI disclosure, with thresholds vague enough that they may add to authors’ uncertainty.
“Generative AI is a powerful resource for health professions education (HPE) researchers publishing their work. However, questions remain about its use and guidance about disclosure is inconsistent.”Europe PMC full text
Empirical study
Cross-disciplinary · Science · 2025-12-18
#Scientific production in the era of large language models
Keigo Kusumegi, Xinyu Yang, Paul Ginsparg, Mathijs de Vaan, Toby Stuart, Yian Yin
Analyzes 2.1 million preprints, 28,000 review reports, and 246 million document views. Detected LLM footprints correlate with higher output and shifted citation patterns under observational designs without causal identification; linguistic complexity shows opposite associations with acceptance and review scores across detected versus undetected manuscripts (using publication outcomes as indirect quality proxies), and Renault et al. later critiqued the first-detection timing design.
“We find: 1) scientists adopting LLMs to draft manuscripts demonstrate a large increase in paper production, ranging from 23.7-89.3% depending on scientific field and author background, 2) LLM use has reversed the relationship between writing complexity and paper quality, leading to an influx of manuscripts that are linguistically complex but substantively underwhelming, and 3) LLM adopters access and cite more diverse prior work, including books and younger, less-cited documents.”arXiv abstract
Live pilot
Cross-disciplinary · Nature Biotechnology / Stanford Agents4Science 2025 · 2025-12-17
#Exploring the use of AI authors and reviewers at Agents4Science
Federico Bianchi, Owen Queen, Nitya Thakkar, Eric Sun, James Zou
Organizes and evaluates Stanford's inaugural Open Conference of AI Agents for Science (Agents4Science 2025), which its organisers describe as the first conference where AI agents served as both authors and reviewers (with humans as co-authors/co-reviewers and public submissions, AI-involvement checklists, reviews and recordings).
“Many fundamental questions are still open: how creative are AI scientist agents? How should human researchers collaborate with them? How capable are LLMs at reviewing scientific work? These questions are difficult to study because journals and conferences currently prohibit AI coauthors and LLM reviewers, and researchers often hide how they use AI.”Primary post
X / LinkedIn
Cross-disciplinary · X · 2025-12-05
#Working-reviewer pushback against journal checkboxes pledging that no AI was used in a review
Ryan Briggs (@ryancbriggs)
Ryan Briggs describes writing his own review, uploading the manuscript to a frontier model with stated non-retention settings, then adding model-identified points he judges useful. He argues that this improves his reviews and objects to blanket no-AI pledges. The workflow includes substantive manuscript review, not merely clarity or tone editing; the post does not establish journal authorization or independently verify privacy safeguards.
“I've noticed when submitting peer reviews that some journals now have a box you have to check that pledges that you didn't use AI when doing the review. This seems dumb to me. I do use AI when I do peer reviews. I do the following: 1. Write my review as I always have”Primary post
Essay / Blog
CS & ML · Pangram Blog · 2025-11-18
#Pangram Predicts 21% of ICLR Reviews are AI-Generated
Bradley Emi
Pangram’s vendor analysis classifies 21% (15,899) of examined ICLR 2026 reviews as fully AI-generated and over half as involving some AI assistance. Detector-labelled groups differ in length and scores. These are detector estimates and observational associations, not verified individual misconduct or proof that AI caused the score differences.
“In order to find out, we analyzed all 19,000 papers and 70,000 reviews from the International Conference on Learning Representations, one of the most important and prestigious AI research publication venues.”Live HTML
Essay / Blog
Cross-disciplinary · The Scholarly Kitchen · 2025-11-06
#Guest Post — AI as Reader, Author, and Reviewer: What Stays Human?
Steve Smith
Reports a Society for Scholarly Publishing seminar panel on AI assistants becoming the first interface that processes scholarly content and summarizes articles off-platform, eroding journal referral traffic and forcing publishers toward content-licensing agreements and machine-readable formats. Separates equity-enhancing language assistance from synthetic manuscript generation while noting that stigma suppresses author disclosure.
“AI assistants are increasingly the first interface that processes scholarly content, summarizing and extracting answers before a person reaches a publisher’s page. Several publishers now report traffic declines from search and referrals.”Live HTML
Preprint
Cross-disciplinary · arXiv · 2025-09-18
#Prestige over merit: An adapted audit of LLM bias in peer review
Anthony Howell, Jieshu Wang, Luyu Du, Julia Melkers, Varshil Shah
A randomized simulation using GPT-4o-mini, ten PNAS papers and four US institutions found lower editor quality scores for low-prestige affiliations across three fields, surviving multiple-testing correction. Revealing identity lowered reviewer rejection recommendations by 1.7 percentage points; the reported ~25% uses the overall mean, not the blinded baseline. A gender-by-prestige interaction survived correction, but individual within-prestige reviewer effects did not. These are model outputs from parallel prompts, not observed journal decisions or a comparison with human reviewers. A separate synthetic-CV exercise used GPT-4.1-nano.
“Revealing author identities lowers reviewer rejection recommendations by roughly 25% of the mean rejection rate despite identical content, indicating that status cues beyond paper quality shape outcomes. Institutional prestige is the dominant cue: papers attributed to low-prestige affiliations receive lower quality scores in every field, a penalty that survives family-wise multiple-testing correction at the editor stage.”arXiv abstract
Essay / Blog
Biomed & Life Sci · Science Integrity Digest · 2025-09-16
#Peer Review Congress Chicago – Day 1
Elisabeth Bik
Conference notes report presentations on self-disclosure and detector-estimated AI use at BMJ, JAMA and AACR, alongside research-integrity investigations. The AACR presentation reportedly found fewer than 25% of detector-flagged authors disclosed use and an association between flagged abstracts and lower progression to review, with quality confounding noted. These are secondary notes on heterogeneous studies, not one validated cross-publisher estimate.
“Disclosure of AI use is not reliable. Less than 25% of authors who used AI (as detected by tool) actually disclosed it. All AI-written Letters to the Editor were rejected without consideration until April 2025.”Live HTML
Forensic / Sleuth
Biomed & Life Sci · Reese Richardson Blog · 2025-09-15
#The world’s laziest peer reviewer
Reese Richardson
Investigative blog documents repeated review templates and citation requests in 104 identified articles as of the post, describing an apparent review mill. A September 23 update reports an archive of over 600,000 review documents from 14 BMC journals; that archive is not a count of fraudulent reviews.
“As of September 23, 2025, I have now downloaded >600K peer review documents for >85K articles at 14 BMC journals.”Live HTML
Preprint
CS & ML · arXiv · 2025-09-12
#Prompt Injection Attacks on LLM Generated Reviews of Scientific Publications
Janis Keuper
Uses 1,000 sampled ICLR 2024 submissions to test simple positive or negative instructions prepended to parsed manuscripts across ten models. Four models gave positive recommendations for 100% of papers under positive injection; baseline recommendation rates and output-format failures varied markedly. Separate parser tests showed hidden text surviving some PDF extraction routes. These are simulated recommendations, not conference acceptances, and results need not generalize to guarded or specialized review systems.
“Our systematic evaluation uses 1k reviews of 2024 ICLR papers generated by a wide range of LLMs shows two distinct results: I) very simple prompt injections are indeed highly effective, reaching up to 100% acceptance scores. II) LLM reviews are generally biased toward acceptance (>95% in many models).”arXiv abstract
Editorial / Comment
Biomed & Life Sci · JAMA · 2025-08-28
#Artificial Intelligence in Peer Review
Roy H. Perlis, Dimitri A. Christakis, Neil M. Bressler, Dost Öngür, Jacob Kendall-Taylor, Annette Flanagin, Kirsten Bibbins-Domingo
JAMA Network editorial (published online 28 August 2025; financial-disclosure correction 4 November 2025) outlining a human-led 'driver-assistance' approach to exploring AI in editorial assessment — naming tasks such as summarization, reporting-guideline checks, and consistency screening as candidates for automation while keeping editors and reviewers in full decision-making control.
“We see our approach as analogous to driver-assistance technologies, beginning with adaptive cruise control or blind spot detection, but when it comes to peer review and editorial decisions, we are not considering fully autonomous driving. Editors and reviewers will not be taking their hands off the wheels or their eyes off the road.”Full-text PDF
Empirical study
Cross-disciplinary · PNAS · 2025-08-04
#The entities enabling scientific fraud at scale are large, resilient, and growing rapidly
Reese A. K. Richardson, Spencer S. Hong, Jennifer A. Byrne, Thomas Stoeger, Luís A. Nunes Amaral
Network analyses and selected case studies identify patterns consistent with coordinated editorial misconduct, image reuse and brokers moving between journals. The observed corpus of suspected paper-mill products grows faster than total publication output, but changing detection effort and field coverage limit inference about population-wide fraud growth. Statistical flags are not individual misconduct verdicts, and the study cannot always distinguish commercial mills from informal networks. It does not establish that generative AI caused the observed growth.
“Our results reveal some of the strategies that enable the entities promoting scientific fraud to evade interventions. Our final analysis suggests that this ability to evade interventions is enabling the number of fraudulent publications to grow at a rate far outpacing that of legitimate science.”Europe PMC
Empirical study
Cross-disciplinary · Nature Human Behaviour · 2025-08-04
#Quantifying large language model usage in scientific papers
Weixin Liang, Yaohui Zhang, Zhengxuan Wu, Haley Lepp, Wenlong Ji, Xuandong Zhao, Hancheng Cao, Sheng Liu, Siyu He, Zhi Huang, Diyi Yang, Christopher Potts, Christopher D. Manning, James Zou
Population-level word-frequency modelling of 1,121,912 papers estimates substantial LLM modification of 22.5% of arXiv computer-science abstract sentences and 19.6% of introduction sentences by September 2024. These are sentence-mixture estimates, not percentages of papers or identified authors. Writing-style and demographic shifts can affect estimates; associations with posting frequency, similarity and length do not establish causality.
“Here we conduct a systematic analysis across 1,121,912 preprints and published papers from January 2020 to September 2024 on arXiv, bioRxiv and Nature portfolio journals, using a population-level framework based on word frequency shifts to estimate the prevalence of LLM-modified content over time. Our findings suggest a steady increase in LLM usage, with the largest and fastest growth estimated for computer science papers (up to 22%).”Live HTML
Essay / Blog
Cross-disciplinary · Substack — AI as Normal Technology (AI Snake Oil) · 2025-07-16
#Could AI slow science?
Sayash Kapoor and Arvind Narayanan
Argues that AI-enabled individual productivity may worsen a production–progress gap when attention, verification and understanding constrain collective progress. Reviews competing explanations and imperfect metrics, and proposes institutional reform and tools for quality control. Explicitly allows beneficial AI uses and says slowdown is not inevitable or uniform across fields.
“Unfortunately, AI companies, science funders, and policy makers all seem oblivious to what the actual bottlenecks to scientific progress are. They are simply trying to accelerate production, which is like adding lanes to a highway when the slowdown is actually caused by a toll booth. It’s sure to make things worse.”Live HTML
Essay / Blog
CS & ML · Pangea Blog · 2025-07-11
#The Hidden Influence: Prompt Injection in ArXiv Research Papers
Joey Melo
Blog post from Pangea, an AI-security vendor, that relays Nikkei Asia’s reporting of concealed review-manipulation prompts in arXiv papers and presents an informal demonstration with ChatGPT, Copilot and Gemini. The responses differ: the author describes ChatGPT as following the injection, while Copilot and Gemini also identify weaknesses. The demonstration illustrates a risk, without a controlled baseline, systematic sample or general model failure rate.
“For context, consider a typical research paper with seemingly harmless white space between lines. Nothing special, right? Well, if you toggle dark mode ON, this white text reveals the secret prompt injection. This isn't an isolated incident; a quick search on arxiv.org for "DO NOT HIGHLIGHT ANY NEGATIVES" uncovers numerous other papers employing the same strategy.”Live HTML
Journal news
CS & ML · Nature · 2025-07-11
#Scientists hide messages in papers to game AI peer review
Elizabeth Gibney
Documents how authors embedded invisible white-text and tiny-font prompts in preprints (Nature independently found 18) instructing AI tools to return positive reviews; researchers quoted in the piece call the practice a form of academic misconduct, or ‘cheating’ that is also a symptom of faulty incentives in academic publishing.
“Researchers have been sneaking secret messages into their papers in an effort to trick artificial intelligence (AI) tools into giving them a positive peer-review report. The Tokyo-based news magazine Nikkei Asia reported on 1 July on the practice, which had previously been discussed on social media. Nature has independently found 18 preprint studies containing such hidden messages, which are usually included as white text and sometimes in an extremely small font that would be invisible to a human but could be picked up as an instruction to an AI reviewer.”Primary post
Empirical study
Biomed & Life Sci · Science Advances · 2025-07-02
#Delving into LLM-assisted writing in biomedical publications through excess vocabulary
Dmitry Kobak, Rita González-Márquez, Emőke-Ágnes Horvát, Jan Lause
Analyzes vocabulary changes in 15.1 million English-language PubMed abstracts from 2010–2024 and estimates a 13.5% lower bound on LLM-assisted abstracts in 2024, conditional on attributing excess style-word use to LLMs. The corpus-level method cannot identify individual AI-written papers or separate direct use from humans adopting model-preferred vocabulary. Differences among subcorpora may also reflect editing practices and publication delays.
“We study vocabulary changes in more than 15 million biomedical abstracts from 2010 to 2024 indexed by PubMed and show how the appearance of LLMs led to an abrupt increase in the frequency of certain style words. This excess word analysis suggests that at least 13.5% of 2024 abstracts were processed with LLMs.”Europe PMC
Editorial / Comment
Cross-disciplinary · Nature · 2025-06-25
#AI, peer review and the human activity of science
Carl T. Bergstrom, Joe Bak-Coleman
Argues that reading literature, drafting manuscripts, and conducting peer review are irreducible cognitive and social practices through which science functions, warning that researchers who cede scientific judgement to LLMs give up agency, skills and the thinking that reading and writing provide.
“When researchers cede their scientific judgement to machines, we lose something important.”Publisher page
Perspective
Cross-disciplinary · Research Ethics · 2025-06-21
#Disclosing generative AI use for writing assistance should be voluntary
Mohammad Hosseini, Bert Gordijn, Gregory E. Kaebnick, Kristi Holmes
Reversing their earlier position, the authors propose voluntary disclosure for revision of existing text for grammar, style and readability, distinguished from generating new prose. They argue that detailed mandates are difficult to enforce and could introduce reviewer bias, and propose optional submission checkboxes disclosed after publication. These are ethical and practical arguments, not measured causal effects of disclosure on bias.
“Nevertheless, as we show in this paper, we have changed our position and now support instituting a voluntary disclosure policy because currently (1) the credit due to machines for assisting researchers is moving below the threshold of requiring recognition; (2) it is impractical (if not impossible) to accurately specify what parts of the text are human-/GenAI-generated; and (3) disclosures could increase biases against non-native speakers of the English language and compromise the integrity of the peer review system.”Europe PMC full text
Institutional policy
Biomed & Life Sci · Astera Institute · 2025-06-02
#Scientific Publishing: Enough is Enough
Seemay Chou
Chou advocates replacing journal-centred publishing and says journal-free requirements already operate at Arcadia Science and Astera, with alternatives funded through Astera and The Navigation Fund. She proposes more frequent open release of data, code and provisional results with public evaluation. Institutional experiences and future proposals are presented, not comparative proof that this system improves research quality.
“I no longer believe that incremental fixes are enough. Science publishing must be built anew. I help oversee billions of dollars in funding across several science and technology organizations. We are expanding our requirement that all scientific work we fund will not go towards traditional journal publications.”Live HTML
Live pilot
CS & ML · ICLR 2025 Workshop Experiment / arXiv · 2025-04-10
#The AI Scientist-v2: Workshop-Level Automated Scientific Discovery via Agentic Tree Search
Yutaro Yamada, Robert Tjarko Lange, Cong Lu, Shengran Hu, Chris Lu, Jakob Foerster, Jeff Clune, David Ha
Presents The AI Scientist-v2 using progressive agentic tree search and VLM figure-refinement loops; in an approved blind test at an ICLR 2025 workshop (the paper notes that workshop acceptance rates are typically 60–80%, and reviewers were told that some submissions might be AI-generated), one of three submitted manuscripts cleared the workshop review threshold (mean score 6.33) before all three were withdrawn under the study protocol. The authors chose the three ideas and picked the best manuscript from several generated for each.
“We evaluated The AI Scientist-v2 by submitting three fully autonomous manuscripts to a peer-reviewed ICLR workshop. Notably, one manuscript achieved high enough scores to exceed the average human acceptance threshold, marking the first instance of a fully AI-generated paper successfully navigating a peer review.”arXiv abstract
Empirical study
CS & ML · ICML 2025 / arXiv · 2025-04-02
#PaperBench: Evaluating AI's Ability to Replicate AI Research
Giulio Starace, Oliver Jaffe, Dane Sherburn, James Aung, Jun Shern Chan, Leon Maksin, Rachel Dias, Evan Mays, Benjamin Kinsella, Wyatt Thompson, Johannes Heidecke, Amelia Glaese, Tejal Patwardhan
PaperBench evaluates replication from scratch of 20 selected ICML 2024 papers using author-assisted rubrics and an imperfect automated judge. Claude 3.5 Sonnet scored 21.0% with BasicAgent; o1 reached 24.4% with IterativeAgent and 26.0% at 36 hours. These are weighted partial-credit scores, not fractions of papers fully replicated. Original code and existing replications were prohibited; human comparisons cover only a small subset and allow AI assistance.
“We evaluate several frontier models on PaperBench, finding that the best-performing tested agent, Claude 3.5 Sonnet (New) with open-source scaffolding, achieves an average replication score of 21.0%. Finally, we recruit top ML PhDs to attempt a subset of PaperBench, finding that models do not yet outperform the human baseline.”arXiv abstract
Journal news
Cross-disciplinary · Nature · 2025-03-26
#AI is transforming peer review — and many scientists are worried
Miryam Naddaf
Charts the dual-track spread of AI across peer review—publishers deploying automated integrity, citation, and statistical screeners alongside reviewers using LLMs, sometimes in violation of journal rules—and captures community concern about leaking confidential information and erosion of the peer-review social contract.
“Artificial intelligence software is increasingly involved in reviewing papers, provoking interest and unease.”Publisher page
Perspective
Cross-disciplinary · Accountability in Research · 2025-03-24
#Disclosing artificial intelligence use in scientific research and publication: When should disclosure be mandatory, optional, or unnecessary?
David B. Resnik, Mohammad Hosseini
Proposes a three-tier framework (online 2025; 2026 issue) making disclosure mandatory for intentional, substantial AI use such as drafting, translation, major revision, data generation, and analysis; optional for grammar, spelling, or organization edits on existing human text; and unnecessary for isolated phrase suggestions or incidental embedded software features, while noting variation across disciplines.
“We distinguish between mandatory, optional, and unnecessary disclosure of AI use, arguing that disclosure should be mandatory only when AI use is intentional and substantial.”Full-text PDF
Journal news
Cross-disciplinary · Nature · 2025-03-07
#AI tools are spotting errors in research papers: inside a growing movement
Elizabeth Gibney
Profiles LLM error-checking initiatives sparked by the black-plastic-spatula calculation mistake (including the Black Spatula Project and YesNoError), weighing rapid detection of errors of fact, calculation, methodology and referencing against false alarms and reputational harm; a clarification of 11 March 2025 added detail on the tools’ current abilities and the concerns about them.
“Late last year, media outlets worldwide warned that black plastic cooking utensils contained worrying levels of cancer-linked flame retardants. The risk was found to be overhyped — a mathematical error in the underlying research suggested a key chemical exceeded the safe limit when in fact it was ten times lower than the limit. Keen-eyed researchers quickly showed that an artificial intelligence (AI) model could have spotted the error in seconds.”Publisher page
Editorial / Comment
Biomed & Life Sci · The Lancet Neurology · 2025-03-01
#Navigating the use of generative AI in peer review
Laura Hart
Sets out the Lancet journals' prohibition (as of March 2025) on generative AI tools in the scientific review of manuscripts, on the grounds that critical thinking and nuanced evaluation cannot be adequately replicated by AI and that uploading unpublished manuscripts would breach confidentiality, while noting that reviewers may have suspected undisclosed AI use in the writing of a manuscript.
“We firmly believe that the critical thinking and nuanced evaluation required for effective peer review cannot be adequately replicated by AI technologies. To that end, the Lancet journals do not allow the use of generative AI tools in the scientific review of manuscripts. Inserting unpublished manuscripts into large language models during peer review would constitute a breach of confidentiality.”Primary post
Perspective
Cross-disciplinary · PNAS · 2025-01-27
#The misalignment of incentives in academic publishing and implications for journal reform
Jennifer S. Trueblood, David B. Allison, Sarahanne M. Field, Ayelet Fishbach, Stefan D. M. Gaillard, Gerd Gigerenzer, William R. Holmes, Stephan Lewandowsky, Dora Matzke, Mary C. Murphy, Sebastian Musslick, Vencislav Popov, Adina L. Roskies, Judith ter Schure, Andrei R. Teodorescu
A perspective on tensions between knowledge dissemination, academic prestige and journal revenue. Discusses modular publishing, micropublications, community-reviewed preprints and changes to academic assessment, including financial sustainability, adoption barriers and unintended incentives. These are proposed reforms and competing arguments, not an empirical demonstration of their effectiveness.
“We also discuss new publishing models and platforms, such as preprint servers, the Peer Community In (PCI) family, modular publishing (e.g., Research Equals), and micropublications.”Europe PMC full text
Community pulse
Cross-disciplinary · Substack — Second Thoughts · 2024-12-18
#The Black Spatula Project
Steve Newman
Launch proposal for the Black Spatula Project, prompted by an o1 demonstration identifying a tenfold arithmetic error in a published study. Proposes small-scale experiments before mass literature auditing and explicitly leaves detection performance, false positives and practical costs unresolved.
“In particular, if this project bears fruit, it will highlight one AI capability that I think is under-appreciated: the ability to grind through an enormous amount of work, such as reviewing a million scientific papers.”Live HTML
Empirical study
Cross-disciplinary · NEJM AI · 2024-07-17
#Can Large Language Models Provide Useful Feedback on Research Papers? A Large-Scale Empirical Analysis
Weixin Liang, Yuhui Zhang, Hancheng Cao, Binglu Wang, Daisy Yi Ding, Xinyu Yang, Kailas Vodrahalli, Siyu He, Daniel Scott Smith, Yian Yin, Daniel A. McFarland, James Zou
Compares GPT-4 feedback with human peer reviews across 3,096 Nature-family papers, 1,709 ICLR submissions, and 666 eLife submissions: point overlap was comparable to human–human reviewer overlap (measuring topical overlap distinct from technical error detection), 57.4% of 308 surveyed authors in a self-selected user study rated the feedback helpful or very helpful, and the authors recommend early-stage author feedback alongside independent human expert review.
“It is important that expert human reviewers deeply engage with the manuscripts and provide independent assessment without relying on LLM feedback.”arXiv abstract
Editorial / Comment
Cross-disciplinary · Research Ethics · 2024-05-11
#Guidance needed for using artificial intelligence to screen journal submissions for misconduct
Mohammad Hosseini, David B. Resnik
Examines ethical and procedural risks as journals deploy AI tools to screen submissions for plagiarism, text generation, and data or image manipulation (online May 2024; 2025 issue; covered by an October 2025 erratum, DOI 10.1177/17470161251378178). Recommends that humans carefully review every case of suspected misconduct flagged by AI, that journals use only well-tested tools, inform authors of flagged irregularities and give them a chance to respond before allegations are forwarded to their institutions, screen all relevant submissions rather than a selection (which could be biased, for example against authors from low- and middle-income countries or low-prestige institutions), and disclose their screening policies.
“Journals and publishers are increasingly using artificial intelligence (AI) to screen submissions for potential misconduct, including plagiarism and data or image manipulation. While using AI can enhance the integrity of published manuscripts, it can also increase the risk of false/unsubstantiated allegations.”Full-text PDF
Empirical study
CS & ML · ICML 2024 / arXiv · 2024-03-11
#Monitoring AI-Modified Content at Scale: A Case Study on the Impact of ChatGPT on AI Conference Peer Reviews
Weixin Liang, Zachary Izzo, Yaohui Zhang, Haley Lepp, Hancheng Cao, Xuandong Zhao, Lingjiao Chen, Haotian Ye, Sheng Liu, Zhi Huang, Daniel A. McFarland, James Y. Zou
A corpus-level mixture model estimates that 6.5%–16.9% of review sentences at four post-ChatGPT AI conferences were substantially generated or modified by LLMs. These are model-based aggregate estimates, not identified reviewers or wholly AI-written reviews; expansion of human-written outlines is compatible with the signal. Higher estimates correlate with deadlines, lower confidence and fewer rebuttal replies without establishing causation. Results depend on historical language distributions and simulated AI-writing practices.
“Our results suggest that between 6.5% and 16.9% of text submitted as peer reviews to these conferences could have been substantially modified by LLMs, i.e. beyond spell-checking or minor writing updates. The circumstances in which generated text occurs offer insight into user behavior: the estimated fraction of LLM-generated text is higher in reviews which report lower confidence, were submitted close to the deadline, and from reviewers who are less likely to respond to author rebuttals.”arXiv abstract