The newsroom-specific case for AI disclosure
Newsrooms sit at the worst point on the AI-disclosure risk curve. Reach is high, the half-life of a trust scandal is years, and an undisclosed synthetic image on a breaking-news story is the textbook fact-pattern regulators want to make examples of. The default reaction in 2025 was to write a one-paragraph "ethics policy" on the "About" page and assume that handled it. By 2026 that approach is failing on three fronts at once: the EU AI Act binds in August, the AP and Reuters standards already require per-asset disclosure, and Google News, Apple News, and the Originator Profile have started reading provenance signals into ranking and inclusion.
The cheapest insurance against any of those outcomes is a per-story label that emits in three places at once: the visible reader-facing banner, the asset-level overlay or metadata, and the machine-readable JSON-LD block in the page <head>. This generator outputs all three for the four AI usage patterns that actually appear in a typical newsroom workflow.
The four AI usages newsrooms actually publish
Not every story needs a label, and overdisclosing dilutes the signal. The generator's presets map to the four cases that genuinely require disclosure under Article 50, AP standards, and the Reuters Trust Principles:
- AI-generated illustrations. The single highest-risk case. Photorealistic AI imagery presented next to a real news event is what the EU AI Act calls out under the deepfake category and what AP's standards forbid outright. The generator's "illustration" preset emits a banner above the headline, a corner overlay sized for a 1200×630 hero, and a JSON-LD
CreativeWorkwithcreatorset to the AI tool andcontentReferenceTimenoting the asset is illustrative, not depictive. - AI-narrated audio articles. The Wapo, FT, Bloomberg, and Axios pattern of converting print stories to audio with an ElevenLabs, Play.ht, or Azure voice. Article 50 names this explicitly as "synthetic audio." The preset emits an audio-pre-roll disclosure and an RSS-aligned
<itunes:summary>note plus aspokenAudioObjectJSON-LD block that names the voice model. - AI translations. A growing share of international editions are machine-translated with light human edit. The generator emits a header line ("Translated by AI from English; reviewed by [editor]") plus an
inLanguage+translationOfWorkJSON-LD pair so the relationship to the original article is preserved for both readers and the IPTC pipeline. - Synthetic broadcast voices. Voiceover for short video clips, breaking-news push notifications read aloud, or a presenter's voice cloned for late-night updates. The generator's strict preset emits a high-contrast banner, an audible pre-roll, and a deepfake-category JSON-LD block, which is what Article 50 specifically requires for cloned likenesses.
Where to place the disclosure on a typical news layout
The legal text says "clearly and distinguishably perceptible at the latest at the time of the first interaction or exposure." For a news article that means three placements working together:
- Above the byline, under the headline. A tight pill at 12–13px on a soft-amber background, one line: "Illustration generated by AI · Edited by R. Chen." Visible before the reader scrolls into the body.
- Inside the lead photo. A bottom-left overlay reading "AI-generated illustration — Not a photograph." This is the placement that survives social-card unfurling, RSS readers, Apple News, and screenshots, because the label is part of the asset.
- In the IPTC/EXIF metadata of the asset. Out of view of casual readers but in plain view of every aggregator, fact-checking pipeline, and content-credentials reader. The generator emits the
photoMetadata.AIcontentMarkerhint and the matchingdigitalSourceTypeNewsML enum value.
Together those three cover the in-page reader, every downstream surface where the asset gets clipped, and every machine-readable consumer of the article.
JSON-LD, IPTC, and C2PA: how the three fit together
JSON-LD is the on-page schema Google and Bing parse and is the cheapest universal signal. IPTC NewsML is what wire services, syndication partners, and most CMS-to-CMS handoffs read; the relevant field is digitalSourceType with values like trainedAlgorithmicMedia for fully synthetic and compositeSynthetic for AI-altered real imagery. C2PA is the cryptographically signed, asset-level standard the Coalition for Content Provenance and Authenticity publishes; full C2PA requires signing at the creation step (Content Credentials in Photoshop, Adobe Firefly, or a camera with C2PA support), which this generator does not do. What it does emit is a sibling JSON-LD block that mirrors C2PA's actions and ai_model fields, so a downstream verifier that doesn't find a signed manifest still finds a structured declaration.
What this is not: an editorial ethics policy
A per-asset disclosure label is the public-facing receipt. It does not replace the internal policy that decides which AI uses are permitted in the first place.
Most newsrooms in 2025 wrote one of those policies. Almost none of them mapped cleanly to a per-story label workflow. The generator presupposes the policy exists — pick the matching preset, fill the editor of record and the model used, paste the output into the article's template. If your newsroom has not yet written that policy, the Poynter Institute's Ethics & Trust resources are the standard reference, and the Columbia Journalism Review tracks newsroom-by-newsroom guidelines as they get published.
Compliance vs. theatre: what bad newsroom disclosure looks like
| Pattern | What it does | Status |
|---|---|---|
| One paragraph in the "Ethics" page | Reader of the article never sees it | Non-compliant |
| Photorealistic AI image with no label | Article 50 deepfake category; AP forbids outright | Aggravated risk |
| Generic "some content may be AI-assisted" site-wide | Doesn't identify the asset; weak under audit | Borderline |
| Visible image overlay only, no schema | Reader sees it; aggregators and AI search don't | Partial |
| Per-asset overlay + per-story banner + JSON-LD + IPTC | Reader, aggregator, fact-checker, regulator all see it | Best practice |
Workflow for a wire-service-shape pipeline
A news CMS does not want a humans-in-the-loop step at publish time. The pattern that works: open the generator once, build the standard variants for the newsroom — "AI illustration," "AI illustration + AI translation," "AI-narrated audio," "Synthetic broadcast voice," "Deepfake/strict" — and store each output as a CMS reusable component. The IPTC digitalSourceType emits at the asset level and rides through the wire. The JSON-LD lives once in the article template's head partial; only the creator, editor, and model fields change per story. Publishing does not slow down; the receipt is in place automatically.
Frequently asked questions
Do news publishers have to disclose AI-generated illustrations and audio under the EU AI Act?
Yes. Article 50 has no traffic floor and the August 2, 2026 deadline is binding. Penalties scale to €15M or 3% global turnover. Newsrooms are an enforcement priority because reach is high and the harm from undisclosed synthetic media is high.
How does the AP, Reuters, and BBC guidance compare to the EU AI Act?
All three already require disclosure and go further than Article 50 on specifics: AP forbids photorealistic AI of real events, Reuters requires editorial sign-off per asset, the BBC requires a named editor of record. The strict preset satisfies all three plus the EU AI Act in one block.
Will an AI label hurt Google News and Discover distribution?
No. Google has signaled provenance is a positive trust factor; what gets demoted is undisclosed scaled AI content. The generator's JSON-LD names the human author, the AI tools, and the editorial review — the exact signal news ranking treats favorably.
Does this support C2PA content credentials and IPTC NewsML metadata?
At the metadata-hint level. C2PA requires cryptographic signing at the creation step, which this generator cannot do. It does emit a JSON-LD block mirroring C2PA actions and ai_model, plus the IPTC digitalSourceType NewsML value and the photoMetadata.AIcontentMarker hint.
How should we label AI translations and AI-narrated audio articles?
For translations: a header line naming the source language, the AI translator, and the human reviewer, plus a translationOfWork JSON-LD pair. For AI audio: an inline pre-roll, an RSS synthetic-voice tag, and a spokenAudioObject block naming the voice model.
What is the highest-risk AI usage in a newsroom?
Photorealistic AI illustrations of real, contested events. That is Article 50's deepfake category and the AP "do not publish" line. The strict preset emits a high-contrast banner, a corner overlay, and a JSON-LD block identifying the asset as synthetic and not depictive.