Do you have to label AI-generated video and voice in the EU?
Yes — and since 2 August 2026 there are two separate duties, not one. Under Article 50 of the EU AI Act, the provider of a generative AI system must mark its output machine-readably so it is detectable as artificially generated. Separately, if you deploy that system to produce a deepfake — image, audio or video that would falsely appear authentic — you must disclose it. Picking a tool that watermarks its exports satisfies the vendor's obligation, not yours. Both duties are live, and breaches carry fines of up to €15 million or 3% of worldwide turnover.
What actually changed on 2 August 2026
Article 50 is the AI Act's transparency chapter, and it applies regardless of whether a system is classed as high-risk. It came into application on 2 August 2026, enforced by national market surveillance authorities. It contains four duties split across two different parties — and that split is the part most creator coverage gets wrong.
| Provision | Who it binds | What it requires |
|---|---|---|
| Article 50(1) | Provider | Tell people they are interacting with an AI system, at first interaction |
| Article 50(2) | Provider | Mark generative output machine-readably, detectable as artificially generated or manipulated |
| Article 50(3) | Deployer | Inform people exposed to emotion-recognition or biometric categorisation systems |
| Article 50(4) | Deployer | Disclose deepfake image, audio or video; also AI-generated text published to inform the public on matters of public interest |
If you are a creator, agency or business making videos with a tool from this directory, you are almost always a deployer. Article 50(2) is your vendor's problem; Article 50(4) is yours. A Veo clip arrives already carrying Google's SynthID watermark — and you can still breach Article 50(4) by publishing it as though it were filmed footage.
The grace period, and why you probably cannot use it
The Digital Omnibus that deferred the Annex III high-risk obligations into 2027 did not defer Article 50. Law-firm briefings published in August 2026 make the same point in near-identical terms: transparency stayed on schedule, complete with fines. One narrow relief survived — generative systems already on the market before 2 August 2026 have until 2 December 2026 to implement Article 50(2) marking. Systems launched on or after 2 August got no grace period, and the deployer deepfake duty was never in scope for it. The extension is a vendor's shield, not yours, and it expires in December anyway.
What counts as a deepfake, and what is exempt
The Act's concern is content that would falsely appear authentic or truthful. That framing does real work: a plainly fantastical scene is a weaker candidate than a photoreal clip of a plausible street corner. Three carve-outs matter here.
- Artistic and satirical work — where output forms part of an evidently artistic, creative, satirical or fictional work, disclosure softens to informing the public in a manner that does not hamper enjoyment of it. A lighter obligation, not an exemption.
- Assistive editing — systems performing only an assistive function for standard editing, without substantially altering the input data, fall outside the marking requirement. Grammar correction is the worked example; whether a filler-word removal pass in Descript lands on this side is what the text does not settle.
- AI text with editorial responsibility — AI text on matters of public interest escapes disclosure where it has had human review and someone holds editorial responsibility.
The unresolved case for this directory is the synthetic presenter. An avatar video from Synthesia or HeyGen that viewers could take for a real person filmed on camera sits much closer to Article 50(4) than a stylised one. That line is not settled, and it is the single point most worth putting to a lawyer.
What the tools in this directory document
This is the buying criterion that did not exist a year ago: does the vendor say, in public, that it marks its output? Below is what CartSignal could find in August 2026. Read the right-hand column — evidence strength varies a lot, and a weak row is a question for the vendor, not proof of absence.
| Tool | What is publicly documented | Evidence |
|---|---|---|
| Google Veo | SynthID imperceptible watermarking across Veo output, plus a visible "veo" mark in the frame. Reporting indicates the visible mark is waived for Google AI Ultra subscribers in Flow. | Official DeepMind material; carve-out from press reporting |
| ElevenLabs | SynthID audio watermarking announced 25 June 2026, starting with free-tier text-to-speech and stated to be expanding to all ElevenLabs audio. Described as surviving trimming, speed changes, metadata stripping and format conversion. C2PA credentials named alongside it. | Official ElevenLabs blog |
| Synthesia | States content is labelled in accordance with Article 50; participates in the Content Authenticity Initiative. Free-tier videos carry a visible watermark. Whether every user video automatically carries a C2PA credential is not stated. | Official Synthesia pages; per-video behaviour unconfirmed |
| Runway | Operates a public Content Credentials verifier for inspecting C2PA manifests. That page documents the inspection tool; it does not state that all Runway output is automatically signed. | Official Runway safety page; embedding unconfirmed |
| HeyGen | Third-party write-ups describe a visible logo on lower tiers plus embedded C2PA metadata. No official HeyGen statement confirming this was found. | Third-party reports only — verify with the vendor |
| Kling AI | Free-tier output carries a visible Kling watermark, removed on paid tiers. No machine-readable marking statement was reachable; Kling's own domains blocked direct retrieval here as in CartSignal's earlier review. | Not documented in reachable sources |
| Resemble AI | Markets watermarking and deepfake detection as core capabilities and publishes its own AI Act watermarking guidance. Confirm scope against your plan. | Vendor marketing and published guidance |
Not documented in this review: Murf, PlayHT, Cartesia, VEED, Pictory, Fliki and Descript — a statement about what CartSignal could find, not a finding that they fail to mark output. If EU distribution matters, ask the vendor and get it in writing.
The two mechanisms vendors actually use
Two standards come up repeatedly, and they are complements rather than rivals. C2PA Content Credentials are signed metadata recording what made a file and how it was edited — rich, but strippable by a re-encode. Watermarks such as SynthID sit in the media signal itself, surviving transformations that destroy metadata but carrying less information. The Commission's Code of Practice on Transparency of AI-Generated Content — finalised 10 June 2026, voluntary, reportedly ~190 signatories by late July — names C2PA as an example of compliant marking and points towards layering both.
The buyer's read: a vendor that only does C2PA is one platform re-encode away from an unmarked file. If provenance matters to you, favour tools that do both.
A practical checklist for creators
- Work out which hat you wear. Making videos with someone else's model makes you a deployer. Shipping a product with a generative feature inside it can make you a provider, with Article 50(2) attached.
- Do not assume the export is marked. Ask the vendor in writing whether marking is on by default, on your plan, and in the export format you actually use.
- Disclose photoreal output yourself. The vendor's mark is invisible to your audience; Article 50(4) is about telling people. A visible on-screen or in-description label is the low-cost move.
- Check your pipeline does not strip the credential. Re-encoding and platform compression drop C2PA metadata. If you rely on it, verify a finished export, not the raw download.
- Diarise 2 December 2026, when the marking relief for pre-existing systems expires.
- This is only the EU layer. Platform rules on YouTube, TikTok and Meta, and laws such as California SB 942, run alongside it with their own definitions.
What this page does not cover
- Whether any specific piece of your content is a deepfake under Article 50(4) — fact-specific, and a question for a lawyer
- The general-purpose AI model obligations in Chapter V, which bind model providers on a separate track
- Whether the marking a vendor describes is actually present in a given export — CartSignal has tested no files
- Any change after 31 August 2026; the Code of Practice's technical annexes are still maturing
FAQ
Do you have to label AI-generated video in the EU?
Since 2 August 2026, two duties apply: the provider marks output machine-readably, and you disclose any deepfake you publish. A tool that marks its output does not discharge your own duty.
Was the watermarking rule delayed by the Digital Omnibus?
Only narrowly. Article 50 stayed in force from 2 August 2026. Generative systems already on the market before that date have until 2 December 2026 for Article 50(2) marking; newly launched systems and the deployer deepfake duty got no grace period.
What are the fines?
Article 99(4) provides for fines up to €15,000,000 or, for an undertaking, 3% of total worldwide annual turnover, whichever is higher. Article 99(6) caps SMEs and start-ups at the lower of the two instead.
Does an AI avatar video count as a deepfake?
It turns on whether the output would falsely appear authentic. A synthetic presenter viewers could take for a real person sits much closer to Article 50(4) than an obviously stylised one.
Which tools already mark their output?
DeepMind documents SynthID across Veo output plus a visible frame mark; ElevenLabs announced SynthID audio watermarking in June 2026; Synthesia states it labels content in line with Article 50. Several other tools here publish no clear statement.
Source links
- Article 50 transparency rules — practical guide
- Article 99: Penalties
- Commission — Code of Practice on Transparency of AI-generated Content
- Commission FAQ — Article 50 transparency obligations
- Goodwin — "Not Delayed, Not Deferred"
- Morgan Lewis — what took effect on 2 August 2026
- Gibson Dunn — Omnibus agreement and postponed deadlines
- Google DeepMind — watermarking video with SynthID
- ElevenLabs — detecting ElevenLabs audio with SynthID
- Synthesia — responsible AI and safety
- Runway — verify Content Credentials