Does the EU AI Act require AI generated content to be labeled?

Yes. Article 50, in application since 2 August 2026, requires providers of AI systems that generate synthetic audio, image, video or text to mark outputs in a machine readable format and detectable as artificially generated or manipulated. Deployers must disclose deepfakes visibly, and AI text published to inform the public on matters of public interest, unless a person holds editorial responsibility after human review.

Is a watermark enough for Article 50?

A watermark alone is a partial answer. Article 50(2) asks for marking that is machine readable and detectable, effective, interoperable, robust and reliable against the state of the art. In practice the defensible stack is layered: signed provenance metadata such as C2PA Content Credentials, an imperceptible watermark for when metadata is stripped, and a visible label where humans must know at a glance.

What is C2PA and why does every regime point at it?

C2PA Content Credentials are signed, tamper evident metadata from the Coalition for Content Provenance and Authenticity that record who made a piece of content and how. They are the common technical denominator the EU's machine readable duty and California's provenance duties can both be satisfied with, which is why the major model providers and camera makers ship them.

Do these duties apply outside the EU and California?

The EU duty follows the market: content reaching people in the EU is covered wherever the provider sits. California's follows users: a generative AI system publicly accessible in California with over one million monthly users is a covered provider wherever it is headquartered. China, South Korea and India have their own labeling regimes with similar mechanics.

Free instrument

Does your AI content carry the marks the law now requires?

On 2 August 2026 two labeling regimes became enforceable on the same day: EU AI Act Article 50 and California's SB 942. Five answers map which duties are yours and whether your current marking stack covers them. Every duty is quoted from the statute, with its source.

Do I have to label AI-generated content?

If you provide a generative AI system to users in the EU, Article 50 of the AI Act requires machine-readable marking of its output, and deployers who publish deepfakes or AI-written public-interest text must disclose it. If you are a covered provider under California SB 942, you owe a latent disclosure in the output, a free detection tool, and a visible label on request. Both sets of duties are enforceable now; the checker below tells you which of them are yours.

1. Which are you?

2. What do you put into the world? (pick every one that fits)

3. Where does it reach people?

4. What does your current marking stack already do?

The convergence, from the texts

Provenance infrastructure stopped being a research topic and became enforceable law on both sides of the Atlantic on the same day. The four layers now stacking up:

EU AI Act Article 50, in application since 2 August 2026

Providers mark synthetic audio, image, video and text as machine readable and detectable; chatbots announce themselves; deployers disclose deepfakes and AI text published to inform the public. Fines up to EUR 15,000,000 or 3 percent of worldwide turnover (Article 99(4)(g)).

California SB 942, operative 2 August 2026 as amended by AB 853

Covered providers (over one million monthly users) embed latent provenance disclosures, offer a visible label option and run a free public detection tool. Penalty $5,000 per violation, each day a discrete violation.

California platform duties, operative 1 January 2027

Large online platforms (over two million monthly users) must detect provenance data that meets an established standard, show it to users, and not strip it. GenAI hosting platforms may not host models that lack the required disclosures.

California capture devices, from 1 January 2028

Cameras, camera phones and recorders first produced for sale in California must embed capture provenance by default, closing the loop from the other direction: real content proves itself rather than fake content confessing.

What the defensible stack looks like

No single technology satisfies every clause, because metadata can be stripped and detectors can be beaten. The stack that answers both statutes layers three things: signed provenance metadata such as C2PA Content Credentials (the machine readable mark Article 50(2) and California's platform duties both read), an imperceptible watermark that survives a screenshot, and a visible disclosure where a person must know at a glance. The major model providers and camera makers are already shipping the first two; the visible layer is the one deployers most often owe themselves.

The EU duties and the fine bands behind them are served in full in the EU AI Act Explorer, Article 50 and the fine calculator. Study aid, not legal advice; the official texts are authoritative.

Questions

Does the EU AI Act require AI generated content to be labeled?: Yes. Article 50, in application since 2 August 2026, requires providers of AI systems that generate synthetic audio, image, video or text to mark outputs in a machine readable format and detectable as artificially generated or manipulated. Deployers must disclose deepfakes visibly, and AI text published to inform the public on matters of public interest, unless a person holds editorial responsibility after human review.
Is a watermark enough for Article 50?: A watermark alone is a partial answer. Article 50(2) asks for marking that is machine readable and detectable, effective, interoperable, robust and reliable against the state of the art. In practice the defensible stack is layered: signed provenance metadata such as C2PA Content Credentials, an imperceptible watermark for when metadata is stripped, and a visible label where humans must know at a glance.
What is C2PA and why does every regime point at it?: C2PA Content Credentials are signed, tamper evident metadata from the Coalition for Content Provenance and Authenticity that record who made a piece of content and how. They are the common technical denominator the EU's machine readable duty and California's provenance duties can both be satisfied with, which is why the major model providers and camera makers ship them.
Do these duties apply outside the EU and California?: The EU duty follows the market: content reaching people in the EU is covered wherever the provider sits. California's follows users: a generative AI system publicly accessible in California with over one million monthly users is a covered provider wherever it is headquartered. China, South Korea and India have their own labeling regimes with similar mechanics.

GAGE briefings tell you which AI regulation deadlines are coming, what they actually require of you, and when a program opens.