SABINE HOSSENFELDER: With AI writing papers at scale, the open question in science is no longer publishability but who is responsible for the content

The Gist

Sabine Hossenfelder says AI can now churn out scientific papers so fast that one professor posted about 200 in nine months, and the main preprint site has had to cap submissions. So the question is no longer whether AI can write a paper that looks publishable. It is who answers for what the paper says, and editors and platforms are starting to hold authors to that. This packet steelmans the argument Sabine Hossenfelder makes in this video, in its best light, for analysis; it is not an endorsement of her conclusions.

Conclusion

Because AI now makes plausible research papers cheap to produce in large numbers, the binding question for scientific publishing has shifted from whether AI-written work can pass as publishable to who is accountable for what it says, and platforms are already responding with limits and author-responsibility rules.

Premises

1. In September 2026 the Washington Post reported that a University of Chicago professor had posted about 200 AI-assisted papers in nine months, and the SSRN preprint platform removed 257 of his papers.
2. From October 1, 2026, arXiv limited submitters to two papers a month, citing a marked increase in dense AI-written papers, and its editorial advisory chair said authors were submitting papers whose contents they likely do not understand.
3. Platforms and editors are shifting the burden to authors: arXiv will ban authors who fail to check AI output, SSRN requires AI disclosure, and a journal editor has stated that authors are ultimately responsible for papers under their name.
4. When text that looks publishable is cheap, appearance no longer filters quality, so accountability for content becomes the constraint.

Assumptions

• The evidence that authors do not understand their own AI-written papers comes from editor interviews (three of seven authors interviewed by a machine-learning journal editor could not answer basic questions) and from arXiv's leadership observing the trend, not from authors' own admissions, and arXiv covers computer science and other fields, not only mathematics and physics.
• Most AI-written submissions described by arXiv are low-value rather than important results; the argument concerns the change in the governing question, not the quality of the papers.

Analysis

Argument type: Inductive.

Potential Fallacies

• Hasty generalization (P1–P3 to Conclusion) — One prominent professor, a few platform policies and a small interview sample are used to characterize a field-wide shift. These are vivid, newsworthy cases, and no base rates are given for how many submissions are AI-written or problematic, or whether a few heavy submitters drive most of the volume. The policy facts are solid, but they support a broad structural claim only loosely.
• False dichotomy (publishability vs. accountability) (Conclusion and P4) — The conclusion frames the issue as a replacement of one question by another. Removals and submission caps are themselves quality and volume filters, which shows that publishability, meaning whether work survives scrutiny, remains live. Verification, reviewer capacity and detection also remain open. Accountability is better described as an added, increasingly salient constraint than as the sole binding one.
• Affirming the consequent / reading policy as proof of diagnosis (P2/P3 supporting the conclusion via P4) — If accountability were the binding constraint, platforms would adopt responsibility rules. But they would also adopt them for moderation workload, liability management or reputation reasons. The policies show that platforms perceive a problem and have chosen an available, enforceable lever. They do not show that accountability is the true constraint.
• Unsupported bridging premise (non sequitur) (P4) — P4 asserts that once appearance stops filtering quality, accountability becomes the constraint. That does not follow automatically. Accountability is an after-the-fact deterrent and does not itself screen content, and it presupposes a way to judge correctness. P4 also assumes that appearance was previously a strong quality filter, when peer review, replication and reputation were the main filters.
• Equivocation on key terms (P1, P4 and Conclusion) — 'Publishable', 'plausible' and 'looks publishable' shift in meaning, as do 'AI-assisted' (P1) and 'AI-written' (P2, conclusion). 'Constraint' (a practical bottleneck) and 'binding question' (a governing normative issue) are treated as the same thing. Holding these terms fixed would sharpen the claim.

Counterarguments

These challenges are not ranked.

• Conclusion (High impact) — The binding question is still quality, now at scale. Cheap plausible papers make verification and attention the scarce resources, not accountability. The cited removals and caps are content and volume filters, because accountability regimes cannot screen work before it floods the system. Publishability, understood as 'does this survive scrutiny', is exactly what AI-scale submission stresses, and the platforms' own evidence supports this reading at least as well as the accountability thesis.
• Premise 4 (High impact) — Appearance was never the main quality filter. Peer review, replication, citation and reputation were, and they still operate. Paper mills, ghostwriting and unreadable-but-formally-correct papers predate generative AI, and author accountability norms were already codified. The change is one of scale and salience, not of kind, so the 'shift' is overstated.
• Premise 3 (Medium impact) — Responsibility rules are weak filters. Bad actors can comply nominally through disclosure checkboxes and signed attestations, or use pseudonymous and throwaway accounts that bans do not deter. Detection of unchecked AI output is unreliable, so enforcement likely depends on after-the-fact complaints. Shifting burden to authors may also be a way for platforms to offload moderation costs.
• Premise 1 (Medium impact) — A single prolific professor is a salient outlier, not evidence of a trend. A heavy-tailed distribution in which a few actors drive most of the volume would justify caps without implying that the central problem of publishing has changed. The 200 versus 257 discrepancy raises additional questions about what was removed and why.
• Premise 2 (Medium impact) — Caps apply to all authors regardless of conduct, so they may penalize legitimate prolific researchers, large collaborations and non-native English speakers who use AI responsibly. The 3-of-7 interview figure is tiny and probably drawn from authors the editor already suspected, and honorary authorship by people who cannot defend a paper predates AI.
• Conclusion (Medium impact) — Putting accountability chiefly on individual authors neglects structural drivers such as publish-or-perish incentives, institutional and funder roles, and AI tool providers. Responsibility is better seen as shared, and incentive reform may be the real lever.

Suggested Improvements

• Scope of the conclusion — Restate the claim as 'accountability has become an additional and increasingly salient constraint alongside quality and volume screening', and qualify 'no longer publishability' accordingly. The platforms' own removals and caps are quality and volume filters, so the weaker claim is supported by the evidence and is not undercut by it.
• Support for P4 — Argue that accountability is the binding constraint and compare it explicitly with reviewer capacity, verification cost and detection. Specify what observation would show which constraint binds. P4 carries the inferential weight and is presently an unargued bridge. Ruling out rivals would turn an interpretive framing into a defended claim.
• Base rates and representativeness — Add data on the share and growth of AI-written submissions on arXiv and SSRN, how concentrated the volume is among a few authors, and how many cases beyond the Chicago professor exist. Without a denominator, vivid cases cannot distinguish a systemic shift from a heavy tail of outliers.
• Reconcile and clarify figures and terms — Explain the ~200 posted versus 257 removed numbers (time windows, paper sets, removal criteria) and distinguish 'AI-assisted' from 'AI-written'. Unexplained inconsistencies and shifting terms weaken trust in the premise and make the evidence look looser than it is.
• Distinguish preprints from journals — State that arXiv and SSRN screen for scope and basic standards rather than correctness, and say what 'publishable' means on each type of venue. A general audience may read preprints as peer-reviewed publications and so misjudge how far the evidence extends to science as a whole.
• Enforcement and effectiveness — Show evidence that author-responsibility rules are enforceable and that they reduce low-quality submissions, and define a threshold for the 'understanding' authors owe. Announced rules do not demonstrate a working mechanism. A vague standard of responsibility could be either impossible or meaningless.

Scenario Tests

• Other major platforms and publishers adopt caps, disclosure mandates and author-ban policies over the next year. (Supports) — This would strengthen the descriptive claim that accountability rules are becoming standard. It would not by itself show that accountability is the binding constraint, because imitation and shared workload pressure also predict convergence.
• Volume rises past what moderators and reviewers can screen, and platforms add content-based triage, endorsement requirements and verification tools beyond author-responsibility rules. (Challenges) — The practical bottleneck would be verification capacity, not accountability. Quality screening would remain central, and accountability would be a supporting rule rather than the governing question.
• AI-generated papers begin to include correct and important results, with authors who understand and verify them. (Challenges) — Quality would return to the center of the question. The argument is time-indexed to a period when most AI submissions are low-value (A2), and its framing could become obsolete as the technology improves.
• Data show a few heavy submitters account for most AI-written volume, and caps plus targeted bans largely end the surge. (Neutral) — The platforms' response would be validated, but the episode would look like a manageable moderation problem and not a change in the field's governing question.
• Bad actors comply nominally by disclosing AI use and signing responsibility attestations while continuing to flood venues, and enforcement proves difficult. (Challenges) — Accountability would be shown to be a norm without a reliable mechanism, so it could not be the binding constraint. Verification-based approaches such as data, code and reproducibility checks would be needed.
• Journals report that authors of AI-assisted submissions routinely cannot answer basic questions about their work, in systematic and non-selected audits. (Supports) — This would give independent support for the claim that authors submit content they do not understand and would strengthen the case that author-level responsibility is the live issue.

Coherence & Relevance

The argument is coherent and plausible at the descriptive level: cheap AI text is producing volume, and platforms are responding with limits and author-responsibility rules. The inference to a replacement of the publishability question by an accountability question is less secure. The factual premises establish that institutions are responding, while the conceptual step depends on P4, which is unargued, and on an unstated factual trigger that plausible papers are now cheap at scale. The stated assumptions helpfully limit the scope (the evidence is observer-based and the claim concerns the governing question, not paper quality). A more defensible version would say that accountability has become an additional, increasingly prominent constraint alongside continuing quality and volume screening.

• In September 2026 the Washington Post reported that a University of Chicago professor had posted about 200 AI-assisted papers in nine months, and the SSRN preprint platform removed 257 of his papers. (Moderate) — It illustrates that scale production is possible and that platforms respond, but it is a single case with no prevalence data. The platform's removals are an act of quality filtering, which cuts partly against the 'appearance no longer filters' framing. The numerical discrepancy is unexplained.
• From October 1, 2026, arXiv limited submitters to two papers a month, citing a marked increase in dense AI-written papers, and its editorial advisory chair said authors were submitting papers whose contents they likely do not understand. (Strong) — It directly supports the claim that platforms are responding to AI volume. It supports the accountability framing less directly, because a cap limits throughput without assigning responsibility, and the 'do not understand' inference rests on small-sample observer evidence.
• Platforms and editors are shifting the burden to authors: arXiv will ban authors who fail to check AI output, SSRN requires AI disclosure, and a journal editor has stated that authors are ultimately responsible for papers under their name. (Strong) — It is the most direct support for the 'responding with author-responsibility rules' clause. It shows response, not that accountability is the binding question, and it does not show that the rules are enforced or effective.
• When text that looks publishable is cheap, appearance no longer filters quality, so accountability for content becomes the constraint. (Moderate) — This is the bridge to the 'binding question has shifted' claim, but it is asserted and not derived. It leaves open rival constraints (verification, reviewer capacity, detection) and does not establish that accountability screens content rather than only sanctioning afterward.

View this argument on LogicFirst.ai