Should journals treat an AI-detector flag as a reason to reject a paper, or just a cue to have a human look closer?

This explores whether a detector flag should start a human review or decide the outcome by itself. The corpus leans toward human review, because the detectors are less reliable than a flag makes them look. ICLR 2026 already works this way: detection tools only sort papers into a queue for area chairs, and an area chair has to find concrete evidence before any sanction is applied How do detection tools shape LLM use enforcement at ICLR?. The trade-off is that wrongly rejected papers become extra reviewer hours. Because a false accusation is so costly for the author, that is the exchange ICLR chose.

The reason that matters shows up when you look at how detectors fail. One study testing rewrites from 23 different LLM versions found no setting that avoids both kinds of error Can one detector catch all LLM versions at once?. Running a separate detector for each version catches almost every AI rewrite but wrongly flags about one in eight human-written abstracts. Training one detector on all versions keeps false flags near 1% but misses about a third of rewrites from the newest models. Under auto-rejection, the first setup would wrongly reject a large share of honest authors. The second would let through exactly the writing the policy is meant to catch. Some rewriting may also slip past detectors entirely, though that claim hasn't been directly tested yet Do rewrites that hide authorship also fool AI detectors?.

The same pattern appears outside AI-text detection. When GPTZero scanned accepted NeurIPS 2025 papers, it flagged hundreds of possibly invented citations. Each one still needed a person to confirm it, and how many were confirmed across the full scan hasn't been disclosed How many accepted conference papers contain hallucinated citations?. A flag tells you where to look, not what you'll find. Work on security monitoring frames the same question as a design choice: if you hold review cost and the false-alert workload fixed, does more monitoring actually improve protection? The paper sets up that comparison but doesn't report results yet Does added monitoring improve protection at acceptable cost?.

There's also a reason to keep humans involved that has nothing to do with accuracy. Peer review problems come from authors, reviewers, and venues together, so a fix that sits entirely with one party, like an automatic rejection rule, misses part of the problem Can two-stage review and badges fix AI conference peer review?. Some misconduct is also only visible to a person reading the paper with intent in mind. Eighteen arXiv manuscripts were found with hidden instructions telling AI reviewers to praise them Are hidden AI prompts in preprints a deceptive research practice?. An area chair can recognise that as deliberate manipulation. A text-style detector isn't looking for it.

The broader point is that a flagging threshold sets who pays for mistakes, as well as how many mistakes are made. Automatic rejection puts the cost of false flags on authors. Sending flags to area chairs puts that cost on reviewers' time. The newest-model numbers also show that any detector will lag behind new models as they appear. Neither choice stops you from tuning the threshold, but with human routing you can change it without wrongly rejecting authors in the meantime.