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
| Situation | Use |
|---|---|
| Launching a campaign with AI-generated video, images, voice or avatars | scripts/disclosure_checker.py on the content manifest |
| Publishing AI-drafted articles, reports or press notes to EU audiences | Decision tree Q4 + checker (public_interest, human_review) |
| Deploying a website chatbot or AI voice agent in the EU | Checker (type: chatbot) + wording library |
| Importing, soliciting or displaying reviews and testimonials | scripts/review_authenticity_linter.py |
| Briefing influencers or virtual influencers | Label templates §6 + Endorsement Guides summary |
| Checking a platform's own AI-label rule | references/platform-ai-label-policies.md |
| Political or electoral advertising | Out of scope beyond platform checkboxes — route to counsel |
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
- 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]
- 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
partialorgenerated. [RECOMMENDED] - Run the checker. Read findings top-down by risk.
FIXlines are unresolved;OKlines are satisfied by the disclosures you declared. - 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]
- Apply disclosures in three layers: legal label on the asset, platform toggle, endorsement disclosure. They stack; none substitutes for another. [RECOMMENDED]
- Record evidence in the asset disclosure record (label text, placement, screenshots, reviewer for the editorial-control exception).
- Re-run until exit 0, then wire the checker into the release pipeline so new assets cannot skip it.
Workflow: reviews and testimonials
- Export the review set you plan to display or quote (own site, marketplace, ad copy).
- Mark
source,incentive,incentive_conditioned_on_sentiment,generated_by_ai,suppressedfor each. - Run
review_authenticity_linter.py. Remove everyfake-or-ai-review,sentiment-conditioned-incentive,review-hijackingitem; add disclosures for insider and incentivised items. - Publish a review-page statement (wording library) saying how reviews are verified — required in the EU (UCPD Art. 7(6)).
- Treat
near-duplicateandposting-burstas investigation leads, not verdicts.
Workflow: EU text on matters of public interest
- 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.
- 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).
- 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)
| Rule ID | Trigger | Default risk | Fix |
|---|---|---|---|
| EU-AIA-50(1) | Chatbot, EU | high | AI-interaction notice at first interaction |
| EU-AIA-50(4)-deepfake | Realistic AI image/audio/video, EU | high | On-asset label |
| EU-AIA-50(4)-deepfake-artistic | Same, evidently artistic/satirical | medium | Non-disruptive disclosure |
| EU-AIA-50(4)-text | AI public-interest text, EU, no editorial control | high | Label or editorial control |
| EU-AIA-50(2)-marking | You provide the generator | medium | Machine-readable marking (legacy systems: 2 Dec 2026) |
| FTC-465.2-fake-review | AI-generated review/testimonial | high, unfixable | Remove |
| FTC-255-material-connection / FTC-465.5-insider | Connection or insider | high | Disclose in the endorsement |
| FTC-255-virtual-influencer | Virtual influencer | medium | "#ad" + virtual persona disclosure |
| NY-synthetic-performer | Paid image/video ad with synthetic performer, NY | high | Conspicuous disclosure |
| EU-UCPD-* / UK-DMCC-* | Fake or undisclosed-connection reviews | high | Remove / disclose |
| IN-IT-Rules-SGI | Realistic AI media, India | medium | Visible label |
| YT / TT / META / GADS | Platform AI-label rules | medium-high | Platform toggle and/or label |
Full summaries with official links: references/regulation-summaries.md.
Key dates and numbers (verify before use)
| Item | Value | Source |
|---|---|---|
| EU AI Act Art. 50 applies | 2 Aug 2026 | eur-lex.europa.eu; digital-strategy.ec.europa.eu |
| Art. 50(2) marking, systems placed on market before 2 Aug 2026 | 2 Dec 2026 | Digital Omnibus on AI; Commission Art. 50 FAQ |
| EU AI Act fine for Art. 50 breaches | up to EUR 15m or 3% of worldwide turnover | Art. 99(4) |
| FTC Reviews Rule effective | 21 Oct 2024 | ftc.gov |
| FTC civil penalty per violation | USD 53,088 (2025 adjustment; see 16 CFR 1.98 for current) | ftc.gov; ecfr.gov |
| UK DMCC fake-review ban | from 6 Apr 2025; fines up to 10% of global turnover | CMA guidance |
| New York synthetic performer ad law | in effect June 2026 | governor.ny.gov |
| India IT Rules SGI amendment | in force 20 Feb 2026 | meity.gov.in |
| Google Ads AI label setting | rolled out July 2026 | support.google.com/adspolicy |
| Meta automated AI detection on ads | from 1 Jun 2026 | about.fb.com |
Exit code contract [PROVEN]
Both scripts share the same contract so they can run as CI gates:
| Code | Meaning | Who fixes it |
|---|---|---|
| 0 | No unresolved findings at or above --fail-on (default high) |
Nobody |
| 1 | Tool error — bad path, malformed JSON, missing id/type/text |
Whoever maintains the manifest |
| 2 | Gate failed — unresolved findings at or above --fail-on |
Asset owner / campaign lead |
Keep 1 and 2 distinct. A malformed manifest is a pipeline problem; a missing label is a content problem.
Anti-Patterns
The Platform Toggle As Legal Compliance
Mistake: Ticking YouTube's or Google's AI toggle and treating the campaign as EU-compliant. Why it happens: The toggle feels official, and the platform shows a label somewhere. Instead: Platform labels can sit in a description or "About this ad" panel. For EU deepfakes, put the label on the asset at first exposure too. Google states its AI label setting does not guarantee regulatory compliance.
The "Composite" Testimonial
Mistake: Asking AI to write testimonials "based on" real survey themes and displaying them with stock names and photos. Why it happens: It feels truthful because the sentiments came from real customers. Instead: A testimonial attributed to a person who did not write it misrepresents the reviewer. Quote real customers verbatim with permission, or present aggregated survey results as statistics.
The Blanket "May Contain AI" Footer
Mistake: One site-wide footer line instead of per-asset labels. Why it happens: It is cheap and feels like cover. Instead: Art. 50(5) requires clear, distinguishable information at first exposure to the specific content. Label the asset; keep the footer as a supplementary policy statement if you like.
The Invisible Editorial Review
Mistake: Relying on the human-review exception for AI-written articles without any record of who reviewed what. Why it happens: Everyone "looked at it" in a shared doc. Instead: Record reviewer, date, substantive changes and the editorial-responsibility holder. Without evidence, the exception is a claim, not a defence.
Incentive Automations That Filter By Stars
Mistake: Sending discount codes only to customers who left 4-5 star ratings, or asking only happy NPS responders to review. Why it happens: Growth tooling makes sentiment-gated flows a checkbox. Instead: Sentiment-conditioned incentives are prohibited by FTC 465.4. Invite every customer (or a random sample) and disclose the incentive.
Troubleshooting
| Symptom | Cause | Fix |
|---|---|---|
| Asset shows "no disclosure rule triggered" but used AI | ai_involvement set to assistive or regions empty |
Re-classify; add every audience region |
| Checker exits 1 | Missing id/type or invalid JSON |
Validate against assets/content_manifest_template.json |
Finding stays FIX after adding a label |
Disclosure key not in needs_one_of |
Use one of the listed keys in disclosures_present |
| Linter misses an obvious disclosure | Disclosure phrased unusually | Put the exact text in disclosure_text; extend DISCLOSURE_RE |
Too many near-duplicate hits on short reviews |
Short texts share phrases | Raise --dup-threshold to 0.7-0.8 |
Scripts
| Script | Purpose |
|---|---|
scripts/disclosure_checker.py |
Content manifest → required disclosures per jurisdiction/platform, risk, gate |
scripts/review_authenticity_linter.py |
Review set → fake/AI, insider, incentive, suppression, hijacking, duplicate and burst findings, gate |
Both: Python 3.8+ standard library only, --format text|json, --fail-on high|medium|low, deterministic.
References
- regulation-summaries.md — EU AI Act Art. 50, FTC 465 and 255, EU UCPD, UK DMCC, New York, India, with official links and penalties
- platform-ai-label-policies.md — YouTube, TikTok, Meta, Google Ads rules as of September 2026
- disclosure-wording-library.md — label text, placement rules, translations, endorsement wording
- disclosure-decision-tree.md — the asset-by-asset decision path and risk levels
Assets
assets/content_manifest_template.json— manifest schema with field guideassets/sample_content_manifest.json— failing sample campaign (exit 2)assets/sample_reviews.json— failing sample review set (exit 2)assets/disclosure-label-templates.md— evidence record, visual tag spec, bylines, chatbot opener, influencer clause