arXiv Implements 1-Year Ban for Papers with Incontrovertible LLM-Generated Errors

What Happened

arXiv is implementing a one-year submission ban for authors whose papers contain what the platform characterizes as "incontrovertible evidence" of unchecked LLM-generated errors. Hallucinated references and fabricated results are the cited triggering examples. The policy targets unverified AI output rather than AI-assisted writing as a category, and arXiv has not published detection methodology, enforcement thresholds, or an appeals mechanism.

Why It Matters

arXiv is the primary pre-publication channel for machine learning and adjacent fields, which means submission eligibility functions as de facto infrastructure for research output. A one-year ban is not a reprimand; it stalls publication timelines, complicates grant reporting, and creates competitive disadvantage against labs that never trigger review. The policy also reframes verification as an authorial obligation rather than a platform function — arXiv is asserting that the burden of citation integrity sits with submitters, and that failure to meet it carries a hard penalty. For operators, this converts citation validation from a soft quality norm into a release-blocking control. The policy likely reduces the volume of low-effort LLM-drafted submissions, but it also raises the cost of legitimate AI-assisted drafting, since every generated claim now carries downstream eligibility risk.

Technical Details

The named failure modes — hallucinated references and fabricated results — are detectable through distinct mechanisms. Reference hallucination is largely tractable via automated checks against Crossref, Semantic Scholar, and DOI resolution, though citation context errors (real paper, wrong claim) require semantic comparison against source text. Fabricated results are harder: they may be internally consistent with a paper's methodology and indistinguishable from error without replication, code inspection, or dataset access. arXiv has not disclosed whether detection is automated, human-reviewed, or triggered by reader reports. "Incontrovertible evidence" implies a high bar, which suggests manual adjudication rather than classifier-based flagging. Absent a stated appeals process, enforcement is effectively unilateral, which creates risk for authors whose errors stem from legitimate ambiguity rather than LLM misuse.

Operational Impact

Research teams using LLM assistance for drafting should treat citation verification as a pre-submission gate, not a post-review cleanup step. Concretely: every reference must resolve to a primary source; every cited claim must be checkable against that source's text; results attributed to external work must be traceable to released code, data, or replication. Teams should log verification steps — tooling, timestamps, reviewer identity — because a provable audit trail is the only defense against an "incontrovertible" finding. Retrieval-augmented drafting pipelines that ground generation in a citation index reduce hallucination risk but do not eliminate it; the final check must be human or a deterministic verifier. Internal review workflows that previously ran in parallel with submission now need to serialize ahead of it.