Is the Answerability Turn Possible?

Nikita Melashchenko · (18 July 2026) · opinion

Luciano Floridi has published the piece I have been waiting for someone to write about AI and publishing. His argument is that the provenance strategy the journals have adopted, detectors, declarations, watermarks and attestations, targets the wrong question. Even a reliable detector would tell us only that a sequence of words came from a particular distribution, not whether the claims are sound or whether anyone stands behind them, and unreliable detectors impose what he calls a “style tax” on non-native English writers whose prose pattern-matches to machines. The function publishing has always tracked, he argues, is answerability, the standing to answer for the claims made, a standing no AI system is eligible to hold. His remedy is a role-indexed architecture in which authors, reviewers, editors and publishers each carry specified answerability, and editorial LLMs sit in the workflow as auditable recommenders that make editors answerable too.^[1]

Earlier I made a narrower, incentives-flavoured version of this point. A profession that shames disclosed AI use, in the patterns Giray catalogues among academic writers, does not get less AI use; it gets undisclosed use, which is exactly the use that skips verification, because admitting to verifying would mean admitting to using.^[2] What Floridi supplies is the standard that diagnosis was missing. AI shaming is incoherent not merely because it backfires, but because it polices the wrong property. It asks whether a machine touched the text, when the only question that ever mattered is who answers for it.

The legal profession should read this carefully, because it is further along the wrong road than publishing. Courts and regulators drafting AI practice notes keep reaching for disclosure obligations and detection, certify whether AI was used, in what tool, for which paragraphs. Yet law already owns the oldest answerability device in existence, the signature on the filing. The lawyer who signs answers for every citation whether it came from a junior, a database or a model, and the disciplinary cases that keep arriving are answerability failures, not provenance failures. Nobody was misled about authorship; someone failed to redeem the promissory notes they signed for. A practice note built on Floridi’s architecture would say less about which tools were involved and more about what the signing lawyer, the supervising partner and the judge relying on machine-assisted summaries must each be able to answer for.

The same holds for legal academia and the classroom. An examiner running student essays through a detector is levying the style tax with extra steps, and a law review demanding AI-free attestations is collecting confessions rather than allocating responsibility. Flip the instrument. Ask the author to state what they answer for, and treat the statement as an assumption of liability rather than an admission of weakness. Disclosure then stops functioning as self-incrimination, which is the mechanism by which shaming drives use underground, and starts functioning as the acceptance of a duty.

The answerability turn, if it happens, will not feel like a revolution. Floridi is explicit that the shift is one of rigour, not principle, and law’s version is older than the printing press. We had the standard all along. It is the signature. And detection was ever a substitute for it.