AI Content Disclosure

AI Content Disclosure

Decide, asset by asset, which disclosures AI-generated or AI-assisted marketing content needs — by jurisdiction and by platform — and block publication when a high-risk item is unresolved. Covers EU AI Act Article 50 (applies from 2 August 2026), the FTC Consumer Reviews and Testimonials Rule (16 CFR Part 465), the FTC Endorsement Guides (16 CFR Part 255), EU and UK fake-review law, New York's synthetic performer ad law, India's synthetic-content labelling rules, and the AI-label policies of YouTube, TikTok, Meta and Google Ads.

Not legal advice. This skill turns public rules into a repeatable pre-publication check. It does not replace counsel, and every rule links its official source so you can verify it. Rules were checked against primary sources in September 2026.

When to use this skill

Clarify First

Before running the check, confirm these inputs. If any is unknown or vague, ASK — do not assume:

• Audience regions — EU, US, New York, UK, India (each switches on different rules; "global" means all of them)
• AI involvement per asset — none / assistive / partial / generated (assistive editing triggers no AI label; generated realistic media usually does)
• Is anything a review, testimonial or endorsement? — and who wrote it, and what they received (fake and undisclosed-connection reviews are the highest-penalty items)
• Paid or organic, and which platform — platform rules differ for ads, and TikTok rejects undisclosed AIGC ads

Stop rule: ask only the 2-3 that most change the output. If the user says "just check it," proceed with the manifest as given and list your assumptions at the top of the report.

Quick start

# 1. Fill the manifest (one entry per asset)
cp assets/content_manifest_template.json my_campaign.json

# 2. Check disclosures — exits 2 if any high-risk item is unresolved
python3 scripts/disclosure_checker.py my_campaign.json
python3 scripts/disclosure_checker.py my_campaign.json --format json --fail-on medium

# 3. Lint reviews/testimonials before display
python3 scripts/review_authenticity_linter.py reviews.json

# 4. Apply wording from references/disclosure-wording-library.md and record it
#    in assets/disclosure-label-templates.md (§1 asset disclosure record)

The shipped samples fail on purpose: sample_content_manifest.json exits 2 with 7 unresolved high-risk findings (a deepfake without an on-asset label, an unreviewed AI press note, AI-written testimonials, an undisclosed employee review); sample_reviews.json exits 2 with 6 high-risk findings.

Core workflow: pre-publication disclosure gate

1. Inventory every asset in the release into the manifest. Include reviews shown in ads and on landing pages, chatbots and voice agents — not just "creative". [PROVEN]
2. Classify AI involvement honestly. "Assistive" means standard editing (spelling, colour, crop, background clean-up). If AI changed what the viewer sees or reads, it is partial or generated. [RECOMMENDED]
3. Run the checker. Read findings top-down by risk. FIX lines are unresolved; OK lines are satisfied by the disclosures you declared.
4. Remove, don't label, what cannot be labelled. AI-written testimonials presented as customer reviews and sentiment-conditioned incentives are prohibited outright — the checker marks them unfixable. [PROVEN]
5. Apply disclosures in three layers: legal label on the asset, platform toggle, endorsement disclosure. They stack; none substitutes for another. [RECOMMENDED]
6. Record evidence in the asset disclosure record (label text, placement, screenshots, reviewer for the editorial-control exception).
7. Re-run until exit 0, then wire the checker into the release pipeline so new assets cannot skip it.

Workflow: reviews and testimonials

1. Export the review set you plan to display or quote (own site, marketplace, ad copy).
2. Mark source, incentive, incentive_conditioned_on_sentiment, generated_by_ai, suppressed for each.
3. Run review_authenticity_linter.py. Remove every fake-or-ai-review, sentiment-conditioned-incentive, review-hijacking item; add disclosures for insider and incentivised items.
4. Publish a review-page statement (wording library) saying how reviews are verified — required in the EU (UCPD Art. 7(6)).
5. Treat near-duplicate and posting-burst as investigation leads, not verdicts.

Workflow: EU text on matters of public interest

1. Decide whether the piece informs the public on a matter of public interest (news-like, policy, health, safety, finance, elections). Product copy normally does not.
2. If yes, choose: label it, or put it under documented editorial control (named reviewer with authority to approve, alter or reject, and an entity holding editorial responsibility).
3. Editorial control is the better default for brands that publish thought leadership — it improves quality and removes the labelling duty. Keep the evidence. [RECOMMENDED]

Rule map (what the checker encodes)

Full summaries with official links: references/regulation-summaries.md.

Key dates and numbers (verify before use)