Do you need to tell people you used AI? Answer three short questions to find out what to check.
Tier 1–4 are this tool’s working categories, not legal classifications. “Not normally” is not a compliance guarantee. Check the source, context and exceptions; optional file tags do not replace public or machine-readable disclosure.
Grouped from lowest concern to highest. Click any row to see what each jurisdiction and platform asks for.
| What was done to the asset | Tier | Tell your audience? |
|---|
Relighting, wardrobe changes, or hair changes on a real person in video. These are normal craft techniques, but they change how a real person looks in a way a background swap does not. Ask whether the change could mislead someone about who is shown or what they really look like. If it could, treat it as Tier 3. If it is clearly styling that nobody would read as a factual claim, it behaves like Tier 2. Do not let it fall into Tier 2 out of habit.
Law and platform policy can impose different duties at the same time. Provider metadata, a public disclosure, a paid-partnership notice and a platform control are separate actions. Choosing the strictest-sounding caption does not necessarily satisfy all of them.
Article 2 defines scope using operator roles, establishment and where outputs are used. A viewer’s location alone does not establish every legal obligation.
For covered deployers, assess three deepfake criteria together: resemblance to people, objects, places, entities or events; a subject that exists or could plausibly exist; and a false appearance of authenticity or truthfulness in context. Realism alone is insufficient.
Public-interest AI text has its own disclosure rule and an exception for human review or editorial control together with editorial responsibility. Creative/fictional deepfakes have adapted disclosure that does not hamper enjoyment. Relevant disclosures must be clear, distinguishable, accessible and provided by first exposure.
Article 50(2) concerns providers’ machine-readable marking of synthetic text, audio, image and video output, subject to its exceptions. This is distinct from deployer disclosure. Article 50 generally applies from 2 August 2026; Article 111’s grace period to 2 December 2026 applies narrowly to Article 50(2) for systems placed on the market before 2 August 2026.
Current Article 4 requires providers and deployers to take measures supporting AI literacy, considering staff and others operating systems on their behalf, their knowledge, use context and affected people. It does not require guaranteeing a specific literacy level. Keeping a note is evidence, not proof the substantive duty is satisfied.
Article 50 breaches fall within Article 99(4)(g): maximum administrative fines can reach €15 million or 3% of worldwide annual turnover. The statute has lower-of treatment for SMEs and, for paragraphs 4/5, qualifying small mid-cap companies. Actual penalties depend on the applicable provisions and case.
Current EU AI Act (Articles 2, 4, 50, 99, 111 and 113)
There is no general federal rule established here requiring every AI-assisted marketing asset to carry an AI label. FTC rules address deceptive claims, endorsements/material connections and fabricated reviews or testimonials. Correct misleading product depictions; a caption does not necessarily cure them.
Fake genuine-experience claims are different from using AI to help a real person word an honest review. Actor, conduct, definitions and knowledge standards in the rule still matter.
FTC review/testimonial rule, 16 CFR Part 465 · FTC advertising and endorsement guidance
GBS 396-b applies to commercial-ad producers/creators with actual knowledge of a qualifying synthetic performer. The definition includes generative AI or software-created representations not recognizable as an identifiable natural performer. Covered ads need conspicuous disclosure; the statute specifies $1,000 for a first violation and $5,000 for subsequent violations.
Important exclusions include audio ads, AI used solely to translate a human performer, certain expressive-work promotions consistent with the work, and transmission media. Do not turn this into “every ad seen in New York must disclose AI.” The law took effect 9 June 2026.
AB 853 made the AI Transparency Act operative 2 August 2026. Its covered-provider definition includes over one million monthly users/visitors and public availability in California. Large-platform/hosting provisions start 1 January 2027; capture-device provisions start 1 January 2028. These are not a universal publisher-label rule.
Future: 1 January 2027. SB 1050 requires specified disclosures in ads prominently featuring qualifying generative-AI synthetic performers, including voices. Its exceptions include certain expressive-work promotions, sole translation and other accessibility features. New York’s broader software-algorithm definition and audio exclusion cannot be copied into this California rule.
State election laws can specify campaign windows, required wording, audio/visual placement, digital metadata, remedies and exceptions. Some laws have been challenged in court. A generic AI caption cannot be assumed sufficient or to make prohibited political content permissible.
Conversational agents, regulated professions, consequential decisions, personal likeness/voice and intimate imagery have additional state-specific rules beyond this asset walkthrough. The state-law report distinguishes full-text checks from search-only screening; neither means “no other law exists.”
Current organic, ordinary-ad and political/social-issue sources were checked. TikTok’s current Community Guidelines were released 25 August 2026 and took effect 24 September 2026. Meta’s ordinary-ad update is dated 1 June 2026. Live help pages may change; account and region controls can differ.
Creator disclosure, automatic labels, advertising disclosures and prohibited uses are separate questions.
Disclose specified realistic or meaningful changes: a real person apparently saying/doing something they did not; altered footage of a real place/event; or a realistic scene that never happened. The headline trigger also covers AI music that is the video’s main focus.
The same page’s examples list synthetic music more broadly. Supporting-soundtrack cases are not clearly settled by that wording; do not assume they are exempt.
Named exceptions include own-voice cloning for a voiceover/dub, production assistance, captions, minor enhancement/repair and clearly unrealistic material. The moving-car backdrop example is not an exemption for all generated backgrounds.
Action: review Attributes → AI use in the upload flow. Some labels are automatic or cannot be changed. YouTube says disclosure alone does not limit audience or monetization; repeated nondisclosure can bring penalties.
Label realistic AI-generated content and significant edits that change meaning, including non-AI digital editing. A false realistic background, added/removed object, person or event can qualify. A clear caption, watermark or sticker can provide disclosure; the platform also offers an AIGC setting.
Minor changes that preserve meaning, artistic/unrealistic styles and generic text-to-speech not recognizable as a known person’s voice have exceptions. A recognizable cloned voice is not automatically covered by the generic-TTS exception.
Action: More options → AI-generated content, where available, or the disclosure method allowed by the current rules. Automatic labels can be irreversible. TikTok says the setting does not affect distribution when content otherwise complies.
Labeling does not permit prohibited private-person likeness use, certain false public-figure endorsements, or harmful false news/crisis depictions.
TikTok current Community Guidelines · TikTok labeling controls
Fully generated image/video/audio and significant AI edits need an AIGC label or clear disclosure. Minor lighting, brightness, saturation, background and denoising edits are listed exceptions. Wholly generated abstract art can therefore have a different outcome in an ad than in an organic post.
Action: non-Spark ads use Ads Manager’s AI-generated-content disclaimer. Spark ads follow the organic post rules. A label does not cure a misleading claim.
Paid political advertising, including compensated creators and Promote, is generally prohibited; there are limited approved government/election-management exceptions. An AI label is not permission to run a political campaign ad.
TikTok paid-ad policy · TikTok ad-disclaimer controls · TikTok political-ad rules
Meta’s current creator-disclosure rule covers digitally created/modified photorealistic video and realistic-sounding audio, including non-AI editing. Examples include realistic AI narration and AI vocals. Cartoon/unrealistic video differs from realistic video.
This mandatory creator rule does not apply to still images. Images can nevertheless receive automatic AI labels from technical signals. AI-generated and AI-edited content can have different label placement.
Action: use the applicable organic AI disclosure control when required, even if an automatic label may appear. Check each content item and current product controls.
The June 2026 update puts AI information in three-dot menu → About this ad → AI info for qualifying own-tool and detectable third-party generative-AI content. Ads may display “Ad” instead of “Sponsored”; rollout and region can differ.
Some regions offer advertiser self-disclosure, including the European Region, California, New York, India and Taiwan. Boosted organic posts may retain organic labels. The source does not establish a global duty to manually label every ordinary AI-written text ad.
Meta current ordinary-ad help · Meta ad-transparency update, 1 June 2026
Where these ads are permitted, advertisers must disclose qualifying realistic digital creations/edits involving apparent acts or speech, people/events that did not exist or happen, or materially altered real events. Minor editing that does not change a claim or issue has exceptions.
Complete the ad’s required digital-content disclosure; automatic labeling is not a substitute. Authorization and disclaimer rules also apply. Meta does not allow political, electoral or social-issue ads in the EU. Nondisclosure can lead to rejection or account restrictions.
Meta political/social-issue ad rules · Meta 2026 election-policy update
Platforms may label content using embedded provenance, their own tools or internal detection. Missing metadata does not prove human authorship. Internal filename tags and automatic labeling do not replace required creator/advertiser disclosures.
C2PA Content Credentials FAQ · Download the full platform audit
The Commission’s icons are optional. Using an icon alone does not establish legal compliance, and the tool’s four working tiers do not determine whether an icon is mandatory.
First assess the actual Article 50 scope, deepfake/public-interest-text criteria and exceptions. A logo, a false testimonial, an abstract image and a synthetic performance do not all have the same legal basis.
Commission icon specifications and license · Current EU AI Act
One synthetic music track or generated person does not prove an entire mixed asset is fully generated. Describe the operation you know occurred; do not infer a percentage of AI use from a medium or filename.
Keep required disclosure clear, distinguishable and accessible by first interaction/exposure. Creative works have an adapted approach that must not hamper enjoyment. The icon guidance permits creative-work exceptions and equivalent user-interface overlays; it is not a universal rule to burn a badge into every file.
Pair ambiguous icons with precise text or audio, provide an accessible description, and allow enough time to perceive the disclosure. A credit or caption is not automatically sufficient for every context.
The Commission uses face-swapped photos and AI-furnished apartments as examples. They do not establish that every object edit requires disclosure or that every product-background swap is exempt. Apply the three deepfake criteria and audience context to your own asset.
The Commission provides black, white and transparency variants. Check its current specifications and license before use.
Reviewed 11 October 2026. Explanations apply the cited general rules; they are not official rulings on your asset.
No. The Commission’s explanation requires three cumulative criteria: resemblance, a subject that exists or could plausibly exist, and a false appearance of authenticity/truthfulness in context. Plausible invented people may qualify, but plausibility alone does not decide the result.
Article 50’s AI duties and platform digital-editing policies have different scopes. TikTok’s significant-edit rule and Meta’s realistic video/audio creator rule can cover non-AI edits. Deceptive advertising, state synthetic-performer rules and rights issues can also apply independently of AI. Manual Photoshop is not a universal exemption.
TikTok current Community Guidelines · Meta current organic labeling · New York GBS 396-b
Article 2 defines relevant roles and territorial connections, including some non-EU operators whose outputs are used in the EU. Merely selecting “EU” here does not establish your legal role or every scope condition. Assess your actual activity and output use rather than an invented follower/view threshold.
Article 50 generally applies from 2 August 2026. The current Article 111(4) grace period to 2 December 2026 concerns Article 50(2) provider marking for systems placed on the market before 2 August 2026. It is not a general delay of deployer disclosure.
No. A filename is an optional internal record. Public disclosure informs the audience. Provider machine-readable marking and platform publishing controls have their own conditions. Material-connection notices serve another purpose. One action need not satisfy all the others.
Assess the actual provider/deployer role, responsibility and control. The Commission explains that contractors operating on a legal person’s behalf under its responsibility/control need not be separate deployers in that situation. “Who decided to use AI” or “who clicked publish” alone is not the complete test. Contracts and platform-account duties are separate.
Article 50(2) provider marking is distinct from Article 50(4) deployer disclosure. Your role matters. Platforms may read technical signals, but automatic labeling is not proof every applicable duty has been met. Preserve provenance where available and complete the required publishing controls.
Article 50 has specific provider standard-editing and other exceptions; deployer creative works have adapted disclosure; public-interest text can have the editorial-review/control plus responsibility exception. Certain authorized law-enforcement uses have exceptions. The conditions differ, so a provider exception cannot be copied into a publisher or platform result.
Current Article 4 requires measures supporting AI literacy in context, without guaranteeing a specific literacy level. Advertising truthfulness, personal rights, licensing and consequential-decision duties may also apply. A training note or AI caption is not a general compliance certificate.
Article 50 breaches are within Article 99(4)(g), with statutory maximums up to €15 million or 3% of worldwide annual turnover. The current statute has different cap treatment for qualifying SMEs and small mid-cap companies in specified paragraphs. Other laws and platform contracts have different remedies; check the actual provision and circumstances.
In the US, purely AI-generated expression lacks human authorship for copyright; sufficiently human-authored expression, arrangement or modification can be protected case by case. Prompts alone do not generally provide the necessary control. Trademark protection concerns source identification, distinctiveness and other requirements. AI use alone neither guarantees registration nor resolves clearance. These are not worldwide ownership conclusions.
US Copyright Office: AI copyrightability findings · USPTO trademark basics
Current statutory text and official platform/help pages are the primary sources. The reports distinguish observed source text from interpretation, inaccessible material and search-only screening. A checked-on date is not a guarantee of unchanged law, enforceability or account-specific controls.
Unsupported specific final-guideline examples, court-denial claims and AI-logo registration counts were removed rather than presented as verified facts.
Current EU AI Act · Commission Article 50 FAQ · Commission icon specifications and license
New York GBS 396-b · Chaptered California AB 853 · Chaptered California SB 1050 · FTC review/testimonial rule, 16 CFR Part 465
YouTube disclosure policy · TikTok current Community Guidelines · TikTok labeling controls · TikTok paid-ad policy · Meta current organic labeling · Meta current ordinary-ad help · Meta political/social-issue ad rules
US Copyright Office: AI copyrightability findings · USPTO trademark basics
The tiers organize a review; they are not statutory classifications. Optional file naming is recordkeeping. Tier 4 is a prompt to examine rights, accuracy or ownership, not a blanket claim that the selected activity is illegal. Seek advice on the specific facts where needed.