EU AI Act Article 50: watermarking and transparency obligations
What the article requires of providers and deployers, what it does not require, and the dates that apply.
What Article 50 requires
Article 50 of the EU AI Act requires providers of AI systems that generate synthetic audio, image, video or text to mark those outputs in a machine-readable format and make them detectable as artificially generated. It separately requires deployers to tell people when they are interacting with an AI system or looking at a deepfake. It does not require the mark to identify anyone.
The obligation sits in Regulation (EU) 2024/1689, and the practical detail of how to meet it sits in the Commission's Code of Practice on Transparency of AI-Generated Content and its Article 50 Guidelines.
The four duties
Article 50 is commonly discussed as though it were one rule. It is four, held by different parties, and confusing them is the most common mistake in the market.
50(1) — tell people they are talking to an AI
A provider must design a system that interacts directly with people so those people are informed they are dealing with an AI, unless that is obvious to a reasonably well-informed, observant and circumspect person.
50(2) — mark generated output
A provider of a system that generates synthetic audio, image, video or text must ensure the output is marked in a machine-readable format and detectable as artificially generated or manipulated. The solution must be effective, interoperable, robust and reliable as far as is technically feasible, accounting for the specificities and limitations of the content type, implementation cost, and the generally acknowledged state of the art.
This is the duty watermarking answers, and the only one of the four a vendor can meet on a provider's behalf.
50(3) — emotion recognition and biometric categorisation
A deployer of an emotion recognition or biometric categorisation system must inform the people exposed to it, and process personal data under the applicable data-protection rules.
50(4) — disclose deepfakes and public-interest text
A deployer who generates or manipulates image, audio or video constituting a deep fake must disclose that the content is artificially generated. There is a modified duty for evidently artistic, creative, satirical or fictional work, and a separate duty for text published to inform the public on matters of public interest — unless that text had human review and a person or organisation holds editorial responsibility.
Provider or deployer — which are you?
Most organisations are one and assume the other's duties. The distinction is defined in Article 3 and decides which paragraphs apply to you.
| Provider | Deployer | |
|---|---|---|
| Who | Develops the system, or has it developed, and places it on the market under its own name or trademark | Uses an AI system under its own authority, other than for personal non-professional activity |
| Duties | 50(1) and 50(2) | 50(3) and 50(4) |
| In practice | A model or product that generates media. Marking happens at generation, in your pipeline | A business publishing or using generated media. Disclosure is something you show a person |
A single company is frequently both — generating media with its own model and publishing it. Those are two separate obligations, and satisfying one does not satisfy the other.
The Code of Practice, and what it expects
Article 50(2) names no technology. The Code of Practice, published in June 2026, is the route most providers are taking, and it describes a layered approach.
| Layer | What it carries | What happens on transformation |
|---|---|---|
| Signed metadata | Provenance assertions attached to the file, typically a C2PA manifest | Removed when a file is re-encoded or re-uploaded without metadata preservation |
| Imperceptible watermark | A machine-readable signal embedded in the pixels, samples or tokens | Persistence depends on the technique and the transformation |
| Fingerprinting | A reference computed from the content, held in a registry | Nothing is embedded; requires a registry to resolve against |
The Code also commits signatories to offering detection of their own marks free of charge to the public, and sets 2 February 2027 for an interoperable detection solution across signatories.
The dates that actually apply
The date most commonly quoted — 2 August 2026 — is right for new systems and wrong for everything already on the market.
- Regulation (EU) 2024/1689 enters into force. Source
- Final Code of Practice on Transparency of AI-Generated Content published. Source
- Article 50 applies. Systems placed on the market from this date must meet the 50(2) marking duty.
- End of the transition for systems already on the market before 2 August 2026 to meet the marking duty.
- Date set in the Code for an interoperable detection solution among signatories. Source
Penalties sit in Article 99: administrative fines of up to €15,000,000 or, for undertakings, up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher. For SMEs and start-ups the lower of the two applies.
Exemptions
The 50(2) marking duty does not apply where:
- the system performs an assistive function for standard editing, or does not substantially alter the input data or its semantics; or
- the system is authorised by law to detect, prevent, investigate or prosecute criminal offences.
The Commission's Guidelines also address content generated in real time that is not recorded and where marking is not technically feasible. That carve-out is narrower than it is usually described: it turns on whether the content is retained and whether marking is genuinely infeasible, not on whether the interaction felt ephemeral. For voice agents in particular, a call that is recorded is not covered by it.
What Article 50 does not require
This section exists because almost every vendor page overstates the requirement — overstating it sells. Being precise about the limits is more useful, and it is what buyers actually ask about.
| Requires | Does not require |
|---|---|
| Marking of synthetic audio, image, video and text in a machine-readable format | Any specific marking technology — the text names none |
| That the output be detectable as artificially generated or manipulated | A visible label applied by the provider; visible disclosure is the deployer's duty under 50(4) |
| Solutions that are effective, interoperable, robust and reliable so far as technically feasible | Identification of the natural or legal person behind the system in the mark |
| Application by the provider of the generating system | Application by the deployer; the deployer holds 50(1) and 50(4) |
The identity point is worth stating plainly: Article 50 does not require a watermark to resolve to a person or a company. Verda offers that, and we think it is valuable — but it is a fraud-prevention and brand-protection argument, not an EU compliance one. Any vendor telling you the EU mandates identity is selling a feature as a mandate.
How Verda meets 50(2)
One watermark, three layers of claim — each mapped to what actually demands it, rather than presented as a single undifferentiated compliance product.
| The duty | Tier | How Verda meets it |
|---|---|---|
| Output marked in a machine-readable format | Compliance | An imperceptible watermark embedded in the signal at generation, across image, audio and video |
| Detectable as artificially generated | Compliance | Free public verification, no account required, plus a documented decode API |
| Effective, interoperable, robust, reliable | Compliance | Signed C2PA manifests alongside the watermark, so the layers back each other up when metadata is stripped |
| Not required by Article 50 — required by California SB 942 | Provenance | The mark resolves to the provider, system, version and timestamp through a registry |
| Not required by any statute | Identity | Optionally the provider record is a KYC-verified organisation or creator — for fraud, brand and agent authorisation |
Verda is registered on the C2PA soft-binding algorithm list for image, audio and video — the public registry the C2PA specification uses to identify watermarking schemes.
Common questions
Does Article 50 require identity in the watermark?
No. It requires that output be machine-readably marked and detectable as artificially generated. Identity is neither named nor implied. California SB 942 is the rule that requires a provider-resolvable payload.
My systems were already on the market. When do I have to mark?
2 December 2026. The 2 August 2026 date applies to systems placed on the market from that date onwards.
Is C2PA enough on its own?
C2PA metadata is one of the layers the Code describes, and it carries the richest provenance while it survives. It is routinely stripped by re-encoding and by re-upload to most platforms, which is why the Code pairs it with an imperceptible watermark rather than treating either as sufficient alone.
Does a visible “AI generated” label satisfy 50(2)?
No. 50(2) is a machine-readable marking duty on the provider. A visible label is what a deployer may owe under 50(4), and the two are not interchangeable.
What are the penalties?
Under Article 99, up to €15,000,000 or 3% of total worldwide annual turnover, whichever is higher, with the lower figure applying to SMEs and start-ups. Fines are imposed by national market surveillance authorities.
Have a question?
Reach out to us at bd@verda.ai