# Legacy example audit — 11 October 2026

Reviewed `work/site/dist/index.html`: jurisdiction comparisons, summary, icons, FAQs, sources/footer, all 28 TIERS rows and 22 ACTIONS entries, and result/template assembly. No website edits made by this reviewer. US state-law and current platform-policy verification belongs to the other reviewers; references below do not independently certify those sections.

## Evidence actually observed

Web opens returned empty bodies. Browser fallback successfully displayed the following primary sources; conclusions below distinguish source text from application/inference.

1. [EUR-Lex consolidated AI Act, 27 July 2026](https://eur-lex.europa.eu/legal-content/EN/TXT/?uri=CELEX%3A02024R1689-20260727): read actual DOM sections Article 4, 50, 99, 111 and 113; the displayed document identifies Regulation 2026/1744 as an amendment. Consolidation itself warns it is a documentation tool; authentic acts are linked from that page.
2. [Commission Article 50 FAQ](https://digital-strategy.ec.europa.eu/en/faqs/transparency-obligations-under-article-50-ai-act): expanded and read deepfake and deployer questions. Last-update date displayed 24 July 2026. Three cumulative deepfake criteria: resemblance, plausible existing subject, and false appearance of authenticity/truthfulness. Context and audience expectations matter; movie backgrounds/effects are a contextual example. Contractors are not separate deployers where acting on the legal person's behalf under its responsibility and control.
3. [Commission EU icon page](https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content): read main content and examples; displayed last update 24 September 2026. Icons optional; use alone does not establish compliance. Four colour/transparency variations; basic/fully-generated/partially-modified variants; face-swap and furnished-apartment examples. Embedding recommendations permit creative-work exceptions and equivalent UI overlays. Licence permits free use without attribution.
4. [Copyright Office AI initiative](https://www.copyright.gov/ai/) and [January 29, 2025 copyrightability findings](https://www.copyright.gov/newsnet/2025/1060.html): human-authored expressive elements, creative arrangements/modifications may be protected; prompts alone are insufficient. AI assistance does not bar copyrightability.
5. [USPTO trademark basics](https://www.uspto.gov/trademarks/basics/what-trademark): read source-identification, goods/services scope and registration discussion. This supports distinguishing trademark from copyright; it does not substantiate a count of registered AI logos.

Article 4 now requires measures supporting staff/others' AI literacy and explicitly does not require guaranteeing a specific individual literacy level. Article 99(4)(g) supplies the Article 50 fine category; lower-of treatment extends to SMCs for paragraphs 4/5 as well as SMEs. Article 111(4) gives pre-August-2-2026 systems a provider Article 50(2) grace period through December 2, 2026. Article 113 retains general August 2, 2026 application and specific other dates. Article 50 distinguishes provider machine-readable marking from deployer image/audio/video deepfake or public-interest-text disclosure; creative works get appropriate limited disclosure, text has editorial-review/control plus responsibility conditions, and disclosure must be clear/distinguishable/accessibly provided by first exposure.

## Claim-by-claim dispositions

The following applications of the verified EU criteria are reviewer inferences, not claims that the Commission ruled on every individual example. Use conditional wording and primary-source links rather than turning examples into categorical legal decisions.

| TIERS example / ACTIONS key | Disposition and smallest revision |
|---|---|
| Dust/scratches/reflections / `cleanup` | Keep as normally low concern. Remove universal exemption: provider standard-editing exception does not independently decide deployer deepfake status; material meaning changes still matter. |
| Exposure/shadows/sharpness/colour | Same. Remove “never enters the labeling question.” Product accuracy and contextual meaning can change. |
| Crop/resize/denoise/format | Same. No universal all-platform guarantee. |
| Upscale/sharpen/repair / `upscale` | Keep routine-editing example, contingent on content/meaning not materially changing. Platform reviewer verifies named exceptions. |
| Own-voice cloning / `voice_own` | EU must assess audio deepfake criteria. Owning the voice is not an Article 50 exception. YouTube's exception cannot be exported to EU law or other platforms. |
| Scripts/outlines/titles/captions | Internal assistance differs from publicly published text. Keep distinction; material commercial claims still require accuracy. |
| Marketing copy / `copy` | Replace “advertising copy is outside” with public-interest purpose + editorial conditions. Health/safety/public-interest information may occur in marketing. |
| Background replacement / `bg_swap` | Conditional. Movie-backdrop example does not establish universal product-photo exemption. Assess whether setting implies a false real-world claim. |
| Seasonal/regional background variants | Same. Location, ingredients/materials and environmental claims may change context. |
| Licensed stock background edit | Same EU test; licensing is separate and not certified by this tool. |
| Unreal VFX / `vfx` | Usually fails false-authenticity criterion when clearly unreal, assessed in context. Remove blanket all-platform exception. |
| Fully generated abstract work / `gen_abstract` | Usually outside deployer deepfake duty, but provider output marking is a distinct question. Plausibility alone does not establish the third criterion. |
| Internal brand exploration | Keep internal/public distinction; do not promise blanket exemption from all AI-system obligations. |
| Invented person in ad / `fake_person` | Plausible invented persons can satisfy existing-subject criterion. Still test resemblance and false-authenticity/context; replace EU universal “Required” with scope/context assessment. State law handled separately. |
| Altered product / `product_alter` | Assess deepfake criteria; deceptive product claim may need correction/withdrawal rather than an AI label. Remove “disclosure is the floor” as universal law. |
| Added/removed objects / `object_add` | Apartment furnishing is verified official icon example, not proof every object edit demands a label. Apply three criteria and context. |
| Product on fake model | Same product-accuracy + synthetic-person assessment. Template may explain illustrative depiction but does not cure material deception. |
| Photorealistic false VFX / `vfx_event` | May constitute deepfake; context matters. Creative-work accommodation is not necessarily no disclosure. |
| False location / `real_place` | Assess contextual authenticity claim. Use medium-neutral template: “Depiction of this location was generated or altered using AI.” Existing image/video action wrongly always says video footage. |
| Colourisation / `colorize` | No blanket EU required conclusion from the statute/FAQ inspected. Apply all three criteria. Unverified specific guideline example needs citation before calling it an official categorical ruling. Remove “legal problem only if fraud/defamation/court” and intent-only promise. |
| AI endorsement / `endorse` | EU can concern realistic AUDIO as well as visual persona. Remove visual-only condition. Material connections are separate; “Paid partnership” must not be inserted without known sponsorship facts. |
| Public-interest text / `public_text` | Keep conditional statutory duty and human-review/editorial-control plus editorial-responsibility exception. Do not infer generation percentage from text modality alone. |
| Generated music / `music` | Original synthetic music alone does not automatically create deployer deepfake duty. A falsely authentic voice/performance may qualify without accompanying visuals. Remove “only unless paired with synthetic visuals.” Provider marking distinct. Platform main-focus condition requires platform reviewer. |
| Real person's replica / `real_person` | May be deepfake; consent/publicity/privacy and other restrictions need context. Remove “every major platform prohibited/restricted” absent current specific sources. Optional icons are not a legal minimum. |
| Altered person's acts / `person_action` | Same. Remove “commonly removed on sight” as unsupported enforcement prediction. |
| False real event / `fake_event` | Scope/context + accuracy issue. No mandatory Fully AI-Generated icon; actual asset may be mixed. |
| Other person's voice / `voice_other` | Assess audio deepfake; ownership/consent/rights separate. Avoid universal consent proposition across every jurisdiction/use. |
| Invented review / `fake_review` | Retain distinction between invented experience and AI assistance wording a genuine experience. FTC-specific actor/conduct scope handled by legal reviewer; “not a labeling option anywhere” exceeds verified scope. |
| Final AI logo / `logo` | Clearly label studio recommendation, not categorical illegality. US AI-only expressive material lacks human authorship, while qualifying human contribution may be protected. No unconditional global copyright claim. Trademark eligibility depends on distinctiveness/use/conflicts. |

## Cross-section inconsistencies to fix together

- Jurisdiction table and FAQ still say audience alone determines EU scope, contradicting corrected walkthrough. Use Article 2's operator/establishment/output-used-in-EU framing; incidental reach is a scope question, not a invented view-count threshold.
- Summary “if no, almost certainly fine,” “always disclose,” “all EU/US/platforms agree,” and “photorealistic/plausible is inside” contradict nuanced result. Replace with three-criteria/context framing and separate law/policy.
- “Whichever rule is strictest” can mislead: different marking/control/disclosure duties can coexist, and one caption need not satisfy all.
- Icon section: remove tier-based mandatory/no-icon statements. Tier 4 includes logos and false reviews; it is not an Article 50 category. Optional icons belong to eligible content when chosen, not an entire tier.
- Icon embedding text must preserve creative-work and equivalent-UI alternatives. “Credits/caption are enough” is too categorical; first-exposure, clear/distinguishable and accessibility duties still matter.
- Apartment example cannot establish a general backdrop/object-addition boundary. Keep it as an example rather than proof of that tier rule.
- Article 4 text must reflect current revised law. “Internal note is enough” confuses evidence with substantive measures. Remove obsolete mere-proposal treatment and unsupported predicted enforcement outcomes.
- Penalties: label Article 50 category, not every AI literacy issue. Include current SMC treatment if describing caps. Avoid “small practice means percentage” without checking turnover.
- Responsibility FAQ needs authority/control/responsibility rather than “who chose AI.”
- Provider marking FAQ should say normally provider role; some organizations can occupy multiple roles. Vendor choice alone does not certify all deployer obligations.
- No general business-to-business exemption established by statute/FAQ read; either cite verified narrow guidance or remove that wording. Client policy is a contract/project requirement, separate from law.
- Copyright FAQ should identify US scope, sufficient human expression case-by-case, and avoid statistics (“hundreds”) without exact primary support. Supreme Court March 2026 denial was not independently checked here; remove or source its court record, and never characterize certiorari denial as a new merits ruling.
- USPTO basics does not prove registered-AI-logo count, AI-specific conflict probability, or “professional search is legally not optional.” Recommend clearance as prudent practice and label studio preference.
- Service Desk provides guidance; do not promise binding legal answers. Footer “ask qualified attorney for a binding answer” also overstates certainty.
- Sources “all links checked August 3” and footer must distinguish historical review, targeted update, verified sources, and remaining gaps. Source presence is not verification of its listed example.

## Generated templates and result UX

- Remove `bg_swap`/`vfx` assurances that subjects/products were photographed/filmed “as they appear”: other selected actions can modify those subjects. State only the selected AI operation.
- `endorse` inserts sponsorship and real/non-real identity claims not established by the selection; use a neutral synthetic-spokesperson statement and a separate material-connection reminder.
- `music` should work for audio-only assets; “This audio includes AI-generated music” avoids incorrectly calling it a video.
- `real_place` must handle photos and videos; neither `_ai-gen` nor partially-modified icon should be inferred solely from this action.
- Multiple captions/icons concatenate conflicting statements. Never designate whole content Fully AI-Generated merely because one music/person element is generated. Explain optional variants and actual composition once.
- Tier 4 studio stop advice should be labelled studio risk guidance. Selecting logo alone must not imply human AI assistance is illegal or universally unprotectable.
- Detailed reasons currently appear only when several actions are selected; a single action also needs the source-backed conditions users must review. Show the one selected explanation without adding a new workflow.
- Preserve three steps. Updated headings/result focus and copy announcements are useful. Keep concise conditional status text with source links beside detailed reasons, and reserve the footer for authorship/history.

## Remaining verification gaps

Specific final-guideline claims about colourisation, sphinx/dragon examples, B2B exception and synthetic influencer were not independently located in the underlying final PDF. Core criteria and contextual movie-background example were directly verified in the Commission FAQ. Supreme Court denial, actual AI-logo registrations and all state/platform enforcement details require the other reviewers or primary records. Mobile layout, screen-reader output and full final-patched decision matrix were not exercised by this reviewer. No website publication performed.