NeurIPS Desk-Rejected 178 Position Papers for Being “AI-Generated.” The Detector Flagged the Track Chairs Too.

1. What actually happened

The NeurIPS 2026 Position Paper Track, chaired by David Rügamer (LMU Munich), Alex X. Lu (Microsoft Research) and Seth Lazar (Johns Hopkins), required all submissions to be “substantially human-written,” permitting generative AI only for “copy-editing or similar peripheral changes”^[1]. That is stricter than the main conference track, where the constraint is on hallucinated citations and undisclosed AI use, not on AI-assisted drafting itself.

To enforce it, the chairs partnered with Pangram, an AI-detection company, and ran every one of the track's 969 submissions through Pangram version 3.3.2^[2]. The tool divided each paper into text windows and assigned an AI-probability score to each segment. Under Pangram's default windowing (approximately 250–350 words per segment), 42.7% of all submissions scored in the 90–100% AI-generated range, and 28.2% — 273 out of 969 papers — received a full 100% score^[2].

The chairs then re-ran the detector using a smaller “medium” window of approximately 100 words, reasoning that narrower windows would reduce over-flagging of papers where AI use was minor or incidental. This dropped the 90–100% flagging rate from 42.7% down to 12.7%^[2]. From the refined results, the chairs split submissions into three buckets:

The 178 desk rejections were not a single threshold. They were driven by three separate criteria^[2]:

That last tier is worth reading twice: 22 papers were desk-rejected for scoring at or above fifty percent on an AI detector, where the author either declared no AI use or left the declaration blank. The gap between the score and the declaration was itself treated as evidence of non-compliance. The chairs describe all three groups as desk rejections “not subject to appeal under standard circumstances.”

The 123 conditional papers were given until June 15, 2026 to submit version histories or document checkpoints showing the state of their paper before and after any AI-assisted edits. If they couldn't, or didn't, the paper was rejected^[2]. Final decisions for the track were released on September 24, 2026^[1]; as of 26 September, the outcome of those 123 cases has still not been published.

2. How they tried to calibrate it, and what happened when others checked

The chairs did attempt validation. Before issuing rejections, they tested Pangram against a “negative control” set: papers from ACM FAccT 2022, a conference whose proceedings predate the widespread availability of ChatGPT^[2]. On that pre-LLM corpus, the detector returned a 0.0% false-positive rate. That is a real, reasonable sanity check, and it is worth giving them credit for running it. The problem is what it doesn't test: whether the detector also returns 0.0% on 2026 position papers that happen to be written in the formal, structured, argumentative style that AI detectors are known to confuse with LLM output.

One researcher ran his own calibration test, and it was more damaging. Sergey Berezin — an AI researcher in correspondence with the track leadership at the time — took recent papers authored by the track chairs and ran them through the same Pangram detector, explicitly as an illustrative check rather than a claim about how those papers were written. The reported AI-probability scores were 24%, 36%, 45% and 69%^[3], and the post spread quickly through the ML community on Reddit and X, where it drew wider discussion of the same distribution-shift problem.

This was not a gotcha for its own sake — it was a calibration test on the exact population the tool was applied to. If the detector cannot reliably clear the professional, peer-reviewed work of the people who chose it, the question is not whether those papers are AI-generated (Berezin himself makes no claim either way), but what that tells you about the false-positive behavior on this population. Recall the three actual criteria above: a 0.69 score doesn't reach 0.9, and it doesn't reach 0.8 either, so neither of the first two tiers applies. The only route by which it meets a desk-rejection threshold is the third one — scoring at or above 0.5 with a declaration of no AI use, or no declaration at all, the criterion that accounted for 22 of the 178 rejections. Had one of the chairs' own papers been submitted with such a declaration, its 0.69 score would have met that tier's threshold. With an honest declaration of some AI use, nothing in the published criteria would have flagged it — which is itself the point: the score alone decides nothing in that tier without the declaration to contradict.

3. The circularity problem

Every NeurIPS submission includes an author attestation. For the Position Paper Track, authors attested their paper was substantially human-written. Here is how the enforcement then worked:

The detector's output functioned as ground truth. If it disagreed with the attestation, the attestation lost. Critics called this circular: the tool's own score was the sole evidence used to contradict the only person who actually knows how the paper was written^[3]. For the 178 outright rejections, the chairs' post describes no stage at which the author could present counter-evidence.

4. What Pangram's own numbers actually say

Pangram 4, released in July 2026, reports an AUROC of 0.9916 and a false-positive rate of 0.0041% — roughly 1 in 24,000^[4]. The same report gives v3.3.2 a false-positive rate of 0.0539% on the same set of over a million human texts. These are vendor-reported figures from the vendor's own benchmarks. Two things make them inapplicable to the NeurIPS decisions:

First, NeurIPS used Pangram 3.3.2, not 4. The improved model was released a month after the desk rejections were already issued. Whatever Pangram 4's accuracy is, it is not the accuracy of the tool that made these decisions.

Second, even if the v4 figures applied, “1 in 24,000” is a rate measured on the vendor's own test distribution. Multiple independent studies have found that AI detectors' real-world performance on formal, structured academic prose — especially from non-native English speakers — diverges substantially from claimed accuracy^[5]. The “simplicity” and formal rigidity of ESL writing statistically resembles LLM output, and a tool trained primarily on informal web text will see patterns that are not there. The track chairs did not publish a calibration study on the specific population of NeurIPS position paper authors, and we could not find one.

5. What the track chairs actually argued

The chairs' published rationale^[2] is more interesting than the backlash suggests. Their central claim was not that AI-generated text is inherently bad, but that it “externalizes the cost of verification onto reviewers” — position papers are supposed to be original arguments, and if the argument was generated by an LLM, the reviewer is now doing the intellectual work the author was supposed to have done. They also raised credit attribution: if a model wrote it, who is the author?

They also published a number that makes the urgency real. Across the NeurIPS 2026 Evaluations and Datasets track, the number of submissions with a Pangram AI score ≥90% increased more than tenfold from 2025 to 2026^[2]. That is not a Position Paper Track problem; that is a conference-wide trend, and the chairs were trying to get ahead of it.

This is, whether or not you agree with how they enforced it, a real argument, and it is the same argument that runs through the hallucinated-citation problem we documented in our last Field Note^[7]. In both cases, an automated system is used to detect something a human reviewer is unlikely to catch at scale, and in both cases the question is whether the detection is reliable enough to stake a desk rejection on.

6. How this compares to the citation checker

This is the second NeurIPS desk-rejection mechanism this cycle, and the two are easy to confuse. They should not be. They check different things, fail differently, and have completely different fairness profiles.

The column that matters most is “Claim type.” A citation checker that says “this DOI does not resolve to any known work in Crossref, OpenAlex, or DBLP” is making a checkable factual claim. You can verify it. An AI detector that says “this paragraph is 73% likely to be AI-generated” is making a statistical claim about style. You cannot verify it by looking at a registry, because there is no registry. One is an audit; the other is an opinion with a confidence interval.

Neither is inherently wrong as a tool. But the standard for using either one to desk-reject a paper with no appeal should be, at minimum, that the tool's error rate on the actual population it is applied to has been measured and published. For hallucinated citations, the independent 48,000-paper audit we cited in our last piece^[6] provides that. For Pangram on NeurIPS position papers, no equivalent study exists.

7. The part that should worry non-native English speakers most

In 2023, a Stanford study tested seven AI detectors on 91 TOEFL essays written by non-native English speakers. On average, 61.22% of those human-written essays were misclassified as AI-generated^[8]. The same detectors were highly accurate on essays by native speakers. The mechanism is structural: non-native writers tend to use simpler vocabulary and more predictable syntax, which is statistically indistinguishable — to a model measuring “perplexity” and “burstiness” — from LLM output.

That study is three years old and did not test Pangram. Pangram's own July 2026 report tested v3.3.2 — the version NeurIPS used — on the same TOEFL essays and reports 0 false positives out of 89, and 2 out of 24,586 across four English-learner corpora^[4]. That is vendor data, on essays and dialogue rather than academic papers, and we found no independent replication. The chairs published no calibration data on non-native authors of formal academic prose, so whether the 2023 bias carries over to this population is unmeasured either way.

One further detail: every submission was shared with Pangram under an enterprise-level data agreement guaranteeing that “zero data from the submissions would be retained” by the vendor^[2]. Whether authors were informed in advance that their unpublished, anonymized research would be sent to a third-party commercial service is not clear from the published materials.

8. If you were rejected: what actually happens to you

This section exists because the people most affected by this are the ones most likely to be reading this article. Here is what we can verify about the consequences:

It is described as a standard desk rejection. The chairs' post calls it “a standard desk-rejection”^[2]; nothing in that post or the call for papers^[1] describes it as a misconduct finding.

You can resubmit elsewhere. Once a paper is no longer under review at NeurIPS, the usual rules for submitting to another venue apply; check that venue's own dual-submission and AI-use policies.

There is no standard appeal channel for the 178. The chairs describe these rejections as not subject to appeal under standard circumstances.

If you were in the 123 conditional group: your deadline to submit version histories was June 15, 2026. Final track decisions were released on September 24, 2026. If you did not submit the evidence, the chairs' post says the paper was desk-rejected.

Practical advice for next time: save every draft. Use tracked changes in your word processor or keep Git commits if you write in LaTeX. If you use any AI tool for anything — even Grammarly, even autocomplete — say so in your attestation. 22 of the 178 rejections were not for high AI scores; they were for the gap between a moderate score and a declaration of no AI use, or no declaration at all. The attestation is not a quiz about what the detector will think. It is a disclosure requirement, and the safest answer is an honest, specific one.

9. What this actually means if you're submitting

This cycle, at NeurIPS alone, authors face two separate automated gatekeepers before a human reviewer ever reads their work. One checks whether your citations are real. The other checks whether your prose is “substantially human-written.” Both can end in a desk rejection. Neither is optional.

The citation-checking problem, at least, has a concrete defense: verify your own bibliography before you submit. Every claim is factual, every disagreement is checkable, and a tool that shows you the specific field-level conflict — this is what you wrote, this is what three registries say — gives you something to act on. That is the thing we built. The content-detection problem is harder, because there is no equivalent self-check: you cannot run your own paper through a proprietary detector whose windowing settings you don't know and whose calibration on your population has not been published. The best concrete advice anyone can give right now is: keep your drafts, save your version history, and if you use any AI tool for anything at all, say so in your attestation rather than claiming zero use — because 22 of the 178 rejections were for the gap between a detector's score and a declaration of no AI use (or no declaration), not for a high score alone.

Check a bibliography for free at strictcite.com. For the other gatekeeper, write your own paper, keep your drafts, and hope the detector agrees.