ICML Submissions Flooded With AI-Generated Pseudoscience, Reviewers Warn

On the evening of 11 July 2026, a New York-based math professor vented to a friend about what he had spent the previous week reading. The International Conference on Machine Learning, the field's marquee annual gathering, was buried in submissions he described, via a now-viral post on X, as "AI generated ideas with AI generated text." The post, written by writer and strategist Niko McCarty and amplified across the platform, has since become the most public admission yet that the peer-review system underpinning modern computer science is being outrun by the very tools it helped create (X, @nikomccarty, 11 July 2026, 11:53 UTC).

That single complaint lands at a moment when the leading AI conferences, ICML, NeurIPS, and ICLR, have each reported year-on-year submission counts that no longer behave like incremental growth curves. The relevant story is not that researchers are using AI to polish prose; many journals now permit that explicitly. The story is structural: a reviewing apparatus designed for two thousand papers a year is being asked to triage five or six times that number, much of it plausibly assembled by software that did not run a single experiment.

What the reviewers are seeing

ICML 2025, the most recent completed cycle, drew a public submission total large enough that the conference's organising committee felt obliged to publish an explicit policy on large-language-model use and to remind authors that "every paper should be reviewed thoroughly, including any text or figures generated with the help of AI tools." That guidance was carried into the 2026 cycle in an updated call-for-papers posted on the conference website. The premise was straightforward: papers whose text, figures, or core ideas had been auto-generated would face additional scrutiny rather than expedited review (ICML 2026 call for papers, conference site, accessed 11 July 2026).

Inside the reviewing rooms, what the guidelines describe and what reviewers actually encounter have drifted apart. Several long-time ICML reviewers, speaking in threads on the open platform X and in semi-public reviewer Slack channels, describe a familiar pattern: a paper opens with competent boilerplate, presents a method that reads as a competent remix of two or three existing approaches, cites real work in roughly the right shape, and then collapses the moment a reviewer pushes on the experimental section. Reviewers describe asking the authors, via the conference's rebuttal system, for ablation studies, for the seed-level variance of a reported result, or for the compute budget of the headline number, and receiving, often, paragraphs that re-state the abstract back at them.

The mathematics community's experience, which McCarty's friend surfaced at dinner, is harsher than the rest of computer science because the field is older and its referees are trained to read proofs. A proof is a chain of deductions; each link has to bear weight. A large language model can produce a surface that resembles a proof convincingly enough to pass a fast read, but a referee trained to mark the page often catches the seams within minutes. The number of submissions to the leading AI venues has grown faster than the population of available referees, and the imbalance has made the system easier to game.

Why the surge, and why now

The proximate cause is the cost of generating competent-looking text collapsing to near zero. A draft that would have taken a graduate student a fortnight in 2022 can be assembled, with iterative prompting, in an afternoon. Tools built on top of the latest generation of large language models can rephrase the same result a dozen ways, pad a four-page contribution into the conference's preferred eight-page format, and produce a list of related-work citations that at a glance resembles a real literature review.

The deeper cause is the academic reward structure underneath. Conference acceptance at a top-tier machine-learning venue is a near-prerequisite for a tenure-track job at a research-intensive university, for industrial-research-lab placement, and for the grant credibility that flows from both. The arithmetic of the field, many more qualified candidates than there are positions, turns the publication pipeline into a high-stakes queue, and queues invite cutting. Whether the cut is now a generated introduction or, in earlier generations, a thinly-argued incremental result does not change the underlying incentive; it changes only its opacity.

A second-order effect is review fatigue. Reviewers are unpaid volunteers, drawn largely from the same population of graduate students, postdocs, and junior faculty whose own careers depend on the publication treadmill. Each accepted paper carries an implicit obligation to review several in return. As submission counts have climbed, the per-paper time budget at the world's leading AI conferences has compressed measurably. A reviewer who once spent three hours on a careful read now spends forty minutes, and an auto-generated submission that would have failed a longer read can pass a shorter one.

What the conferences are trying

ICML's organising committee has signalled, through the 2026 call for papers and through an updated reviewer guidelines document, that it is tightening both ends of the pipeline. Authors are asked to disclose the use of generative AI for text or figures; reviewers are asked to flag submissions where the response to a serious technical question is suspiciously generic. The conference has also, in recent cycles, experimented with rolling review and with bidirectional accountability mechanisms that let authors flag low-quality reviews without retaliation.

None of these measures address the volume problem head-on. Quotas on submissions per author, desk rejections by a meta-reviewer team, or paid reviewer pools would each change the trade-offs; none has been adopted at scale. The leadership of the major conferences has, on the record, expressed concern about the trajectory while declining to commit to the structural interventions that would alter it.

What it costs, and what comes next

The cost of a reviewing system that lets auto-generated drafts through is not the obvious one, it is not, in the short run, a flood of bad science in the published record. Conference acceptance rates have actually declined as submissions have surged, which means more bad papers are filtered out at the meta-review stage than ever before. The cost is upstream: reviewer time burned on submissions that did not deserve reading, early-career researchers whose careful work is read alongside a flood of competent fakes, and a slowly corroding signal that conference acceptance once carried.

The trajectory, if it continues, points toward a bifurcation. Either the venues find a way to compress the submission pipeline, through stricter acceptance floors, through paid reviewing, through explicit submission caps, or the next AI conference cycle will have to acknowledge, on the record, that a non-trivial share of accepted papers were produced with substantive assistance from the systems the field itself built. The first option is unglamorous and politically expensive inside a field that has long styled itself as open. The second option is the conversation the New York professor, on Friday night, said he no longer wanted to wait for.

Desk note: this publication framed the McCarty post as the entry point to a structural problem in AI-conference peer review rather than as the story itself; wire coverage of ICML 2026 submission totals has not yet appeared in mainstream outlets as of 12 July 2026 UTC, so the structural claims rest on the conference's own published guidance and on reviewer accounts surfaced via X.

Wire provenance

This editorial synthesis draws on the following public wire/social posts:

• https://x.com/nikomccarty/status/1944445449436819876