Labeling AI-generated images and video: what Article 50(4) requires since August 2
Two duties now touch every AI-generated image and video an agency publishes. Since 2 August 2026, Article 50(4) of the EU AI Act (Regulation (EU) 2024/1689) obliges deployers to disclose AI-generated or manipulated image, audio, and video content that constitutes a deep fake. And since the same date, the countdown runs toward 2 December 2026, when the separate machine-readable marking duty in Article 50(2) comes fully into force for providers. If your agency produces AI visuals for clients, both dates shape your production pipeline — one is already binding, the other lands before year-end.
The two duties, separated cleanly
Most confusion online comes from merging these into one "labeling rule." They bind different actors and have different dates:
| Disclosure of AI content — Art. 50(4) | Machine-readable marking — Art. 50(2) | |
|---|---|---|
| Who | Deployers — anyone publishing or circulating the content under their own authority. Agencies included. | Providers of AI systems generating synthetic audio, image, video, or text output — the tool vendors, including general-purpose AI vendors. |
| What | Tell people the content was artificially generated or manipulated, clearly and distinguishably, at latest at time of first exposure (timing rule in Art. 50(5)). | Mark outputs in a machine-readable format and detectable as artificially generated — effective, interoperable, robust and reliable as far as technically feasible, per the state of the art reflected in technical standards. |
| Covers | Content constituting a deep fake — generated or manipulated image/audio/video resembling real people, objects, places, entities or events and capable of falsely appearing authentic. Text published to inform the public on matters of public interest has a parallel disclosure duty, with an exemption where human editorial review and editorial responsibility exist. | Synthetic outputs generally — with a carve-out where tools perform an assistive function for standard editing or don't substantially alter input data or its semantics (background removal, noise reduction, straightening). |
| Since / until | Live since 2 Aug 2026. Artistic, satirical, fictional works get a modified version: disclose the existence of generated content appropriately, without hampering display or enjoyment of the work. | Transitional period ends 2 Dec 2026 — from that date, unmarked synthetic output is a straightforward compliance gap for tool vendors. |
Why agencies should care about the second column even though it binds vendors: your choice of generation tools after 2 December 2026 becomes part of your own story. Clients, platforms, and eventually authorities will ask whether the tools in your stack mark their output — and whether your export pipeline preserves those marks.
Platform by platform: how labeling works in practice today
The Act sets the duty; platforms are where disclosure actually happens. Practices below reflect each platform's published documentation — verify current behavior before relying on any toggle, since these interfaces change frequently.
METAFacebook & Instagram
- Auto-labels detected AI content as “AI Info.” Meta began applying labels to a wider range of video, audio, and image content from May 2024, based on industry detection signals — labels now display as "AI Info" (Meta Transparency Center).
- Detection leans on invisible markers: metadata from other companies' tools and standards such as C2PA Content Credentials and the IPTC "Digital Source Type" field can trigger the label — meaning your upstream pipeline choices determine whether Meta can see what the content is.
- Self-disclosure matters most for ads and branded content: ad systems require advertisers to disclose realistic AI-made or digitally-altered media; leaving disclosure to auto-detection is riskier than declaring up front.
- Practical step: declare during upload/ad setup whenever content is realistically AI-generated, and preserve provenance metadata through your export pipeline so platform detectors aren't blind.
TTTikTok
- Creator toggle: before posting, enable the "AI-generated content" switch (Post settings on photo/video posts). TikTok then applies a creator disclosure label (TikTok Help Center).
- Auto-labeling too: TikTok may apply its "AI-generated" label automatically when it identifies fully generated or significantly AI-edited content — including when uploads carry Content Credentials (C2PA) attached, or when TikTok's own AI effects were used.
- Realistic-only rule: TikTok's requirement targets realistic AI content; clearly animated or unrealistic content is treated differently — but the toggle is cheap insurance either way.
- Practical step: make the toggle a mandatory checkbox in your social-team SOP for anything with generative components; don't rely on auto-detection catching your export.
YTYouTube
- "Altered content" disclosure in YouTube Studio: creators must disclose realistic content made with altered or synthetic media via the "AI use" setting during upload; viewers see a label (YouTube Help; YouTube Blog). Announced requirement dates to March 2024, well before the AI Act deadline — so client channels should already comply.
- Scope is "realistic": clearly unrealistic, animated, or cosmetic adjustments don't need the main-stage label; photorealistic scenes, synthetic voices resembling real people, and altered footage do.
- Platform-side additions: YouTube has been extending its own labeling, including automatic labels for some undisclosed realistic synthetic content.
- Practical step: add "altered content? yes/no" to your upload checklist; treat synthetic voiceover as disclosure-triggering even when the visuals are real footage.
Your own site and client properties
No platform toggle exists for assets you serve yourself. Here, Art. 50(4) disclosure is yours to build:
- Caption or on-image marks: a visible caption ("Illustration generated with AI"), a corner badge baked into hero images, or a hover/sr-only annotation — the duty requires the information to be clear and distinguishable at first exposure (Art. 50(5)), so a tooltip buried three interactions deep is a weak answer.
- IPTC Digital Source Type metadata: embed the industry-standard field (values like
trainedAlgorithmicMediafor fully generated,compositeWithTrainedAlgorithmicMediafor edited composites) directly in the asset — this survives into DAMs and feeds platform detectors downstream. - C2PA Content Credentials: where your generation or editing tools attach them, preserve them through export and CDN steps. Stripping provenance in post-production undoes the machine-readable trail that both platforms and the Art. 50(2) regime expect to rely on.
- A published AI-use note: a short page describing how the site uses generative media complements, but does not replace, per-asset disclosure.
The December 2 machine-marking deadline, explained without jargon
Article 50(2) asks something different from the visible labels above: outputs must carry a machine-readable signal — data a computer can read to establish "this was generated or manipulated" — and that signal should be detectable, interoperable, robust and reliable as far as technically feasible. Think watermark-like signals and signed provenance metadata rather than human-facing badges.
- Who must act: providers of generative AI systems — including general-purpose AI vendors whose systems produce text, images, audio, or video. Not the agency. But see below.
- The date logic: the obligation applies from 2 August 2026 like the rest of Art. 50, with a transitional period that ends 2 December 2026 for this specific marking duty — the last milestone on the 2026 calendar.
- What changes for agencies: vendor selection questions become compliance questions. Before renewals: "Does your system mark outputs in a machine-readable format?" and "Do your exports preserve those marks?" After 2 December 2026, a vendor who answers "no" is handing you unmarked synthetic content — and your published pipeline inherits the awkwardness of explaining that choice.
- Status of standards: the Act points to the generally acknowledged state of the art as reflected in relevant technical standards, and the AI Office is charged with facilitating codes of practice on detection and labelling (Art. 50(7)). Expect detail to keep evolving; document whatever your stack does today.
Three worked examples
Example 1 — Hero image for a client landing page
An agency generates a photorealistic "office team" image with a diffusion model for a consulting client's homepage. No real people or places are depicted, but it reads as photography.
Handling: caption or badge disclosing AI generation at first view (Art. 50(4)/(5)); embed IPTC Digital Source Type trainedAlgorithmicMedia; log the asset in the labeling register. Even where a strict deep-fake reading is arguable — nothing real is being impersonated — disclosure is the low-cost, defensible default, and several platforms will auto-label it anyway.
Example 2 — Product visual with a fabricated background scene
A real product photo is composited onto an AI-generated mountain landscape for a social campaign.
Handling: the composite fabricates scenery around a real product — use the platform disclosure toggle (Meta declaration, TikTok toggle, YouTube altered-content setting depending on destination), keep the provenance metadata intact, and disclose in-copy where posted natively ("Scene created with AI"). Assistive-editing exemptions don't cover adding a whole generated environment.
Example 3 — AI video avatar of the client's founder
A synthetic presenter resembling the founder delivers a script for an explainer video.
Handling: this is squarely in deep-fake territory — a generated video of a real, identifiable person. On-screen or verbal disclosure at the start ("This video features an AI-generated presenter"), platform disclosure settings enabled, and written client sign-off documenting who authorized the likeness use — which is additionally a personality-rights question worth a lawyer's glance.
The 60-second agency workflow
- Create: prefer tools that attach provenance (C2PA/IPTC) and ask vendors about Art. 50(2) marking plans before renewals.
- Export: configure pipelines to preserve metadata end-to-end; test one file monthly.
- Publish: platform toggle on where applicable; visible caption or badge on owned properties; note it in the asset record.
- Log: asset URL, tool used, disclosure method, date. Four columns — that's the register.
Keep reading
- Is your website chatbot legal? — the Art. 50(1) disclosure checklist with paste-ready lines.
- Every EU AI Act deadline agencies actually face — the full post-Omnibus table, corrected against primary sources.
Not sure how much of your media stack is exposed?
The free OX-00 self-score walks through disclosure, marking readiness, literacy, and governance in ten questions — mapped to the August and December dates above. Two minutes, no email wall.
Take the free AI-readiness self-score →
Want the gaps turned into a prioritized plan? The paid AI Act Ready packet includes a labeling-log template and ships within 72 hours.