AI text detectors are good enough to deploy. The appeals process is what nobody has written.

The arithmetic in the marketing is worth doing before anyone drafts a policy. Pangram's website advertises 99.98% accuracy ^[4], which permits one error for every 5,000 documents screened ^[1]. GPTZero advertises 99% ^[5], which permits one in a hundred, fifty times as many ^[2]. Two products in the same market are not separated by a factor of fifty in quality. They are separated by what was measured and on whose text. Both figures are the firms' own, and Nature's reporting puts the independent view more narrowly: the tools call human-only text human almost all the time ^[10], and a flag is a starting point for investigation rather than a finding ^[11].

The AACR case shows what resolution looks like in practice. Daniel Evanko asked, and the reviewer confessed to using a model after running out of time ^[2]. He did not expect that, because most researchers do not disclose AI help ^[3]. Take the confession away and the editor is holding a probability and a denial, which is the part of the workflow nobody has designed. The person being scored cannot inspect the system doing the scoring, and proving you wrote your own sentences is not a task with a defined end.

The disputes will not mostly involve wholly synthetic text. Pangram's reading, in a study reported in January, put some AI-generated text in one in eight biomedical articles last year ^[7], and the tools are weakest precisely where that volume sits: writing in which human and machine prose intertwine, with no clean boundary for what counts as problematic use ^[12]. Tim Requarth of NYU Langone says the detectors are good for screening out places pumping out slop ^[13], and that case is easy. Deciding whether a model that tightened a reviewer's grammar breached a disclosure rule is a definitional question, and a score does not answer it.

Opting out is not on offer either. Pangram verdicts sit on arXiv papers through the alphaXiv mirror ^[8], on Substack posts since July ^[18], and on coursework at the University of Chicago ^[9]. Its co-founder Max Spero has publicly flagged journalists and, on one occasion, the Pope's social-media posts ^[17]. Complaints will therefore reach research-integrity offices from third parties whether or not the office screens anything itself. The accused will have a ready defence in the record: before this generation of tools, detection software was notoriously unreliable, in the account of Simon Fraser University computer scientist Marzena Karpinska ^[14], because the statistical approaches flagged human-authored text as AI too often ^[15]. An office that wants its flags to carry weight has to publish its threshold and its appeals route before it uses one in anger.