Article 50 of the EU AI Act
Is your AI-generated image legally disclosed?
Article 50 is already in force. Most teams don't know whether an AI image needs a visible disclosure, or whether the one it has is enough. signator.ai checks a single image against C2PA provenance, AI-detection, and the Act's actual rules — and tells you exactly what's missing.
No account needed · Free · Under a minute
Images today — video and audio are nextAug 2, 2026
Article 50 in force
€15M / 3%
max fine, or of global turnover
2
separate duties, different parties
Providers — the companies building the AI system — must mark generative outputs with machine-readable, detectable signals. Deployers — anyone using AI professionally, including agencies and publishers — must clearly disclose deepfakes and certain AI-generated text to the people who see it.
Critically: a provider's embedded mark does not, by itself, satisfy a deployer's disclosure duty. They're separate obligations, on separate parties. And exemptions are narrower than they sound — satire and fiction get an adapted form of disclosure, not a waiver, and the human-editorial-review exemption applies only to text, never to images.
Three separate findings. Never one blended score.
A detector's confidence and a legal judgment aren't the same kind of probability — collapsing them into one number would be indefensible.
Provenance
Is a C2PA manifest present, is the signature valid, who signed it. A deterministic, cryptographic check — no AI involved.
Detection
Hive's AI-generated and deepfake confidence score, always framed as a probability estimate — never stated as a certainty.
Disclosure
A decision tree reads the Act against your context and states what's required — or says plainly when it can't determine yet.
See what you get
A real example report — not a mockup of the interface, the actual output shape.
Where this is going
signator.ai is a free public beta today — everyone gets full access, including saved reports, no paywall. What's next:
Saved, versioned evidence trail
A repeatable workflow for agencies reviewing AI-assisted campaign assets — each report dated and tied to the exact decision-tree version that produced it.
Video and audio checks
Images only for now. Video and audio detection cost far more per check and aren't needed to validate the idea yet.
Team dashboards and batch checking
Reviewing many assets at once, and letting a team see each other's checks in one place.
Re-checking published versions
Catching drift between the asset that was approved and what actually went live — not just the original upload.
A paid agency tier arrives once real usage and direct conversations with agencies confirm it's worth paying for — not on a fixed timeline.
Want to know when video and audio support ships?
Only used to email you about major updates — no spam, no sharing. See our Privacy Policy.
Try it yourself
Upload an image and answer a few questions. Free, anonymous, no account needed.
Frequently asked questions
The parts of Article 50 people usually ask about first.
Am I a "provider" or a "deployer" under the EU AI Act?+
Providers are the companies building or offering the AI system itself. Deployers are anyone using AI professionally — agencies, publishers, marketers, businesses — to create or publish content. If you're generating or editing images with an existing AI tool rather than building one, you're almost always a deployer, and it's the deployer's disclosure duty that applies to you.
Does an embedded C2PA mark or watermark satisfy my disclosure duty?+
No. A provider's machine-readable mark and a deployer's disclosure duty are separate obligations on separate parties under Article 50. Even if the AI tool you used embeds a valid C2PA manifest, that does not by itself satisfy your own duty as a deployer to clearly disclose the content to the people who see it.
What actually counts as a "deepfake" under Article 50?+
Article 3(60) sets a two-part test: the content has to resemble a real, identifiable person, place, or event, AND it has to be able to reasonably appear authentic or truthful to a viewer. Both conditions need to be true — an obviously stylized or fantastical AI image usually won't meet this bar, but a realistic AI photo of a real-looking scene or person likely will.
Is satire or fiction exempt from disclosure?+
No — this is a common misconception. Satirical and fictional work isn't exempt; it gets an adapted form of disclosure (for example, unobtrusive labeling, credits, or a description) that doesn't interrupt the work, rather than a full waiver of the obligation.
Does editorial review let me skip disclosure?+
Only in a narrow case: the exemption applies exclusively to text, only on matters of public interest, and only where there's genuine human editorial review with a named responsible editor. It never extends to images — an AI image accompanying an edited article still needs its own assessment.
What happens if I don't disclose when I'm required to?+
Penalties under the EU AI Act can reach €15 million or 3% of global annual turnover, whichever is higher. Article 50 has been in force since August 2, 2026.
What if I'm not sure how the image was made, or whether it needs disclosure?+
signator.ai's decision tree returns "cannot-determine" rather than guessing whenever the context it's given isn't enough to reach a real answer. That's treated as a valid, expected outcome — a wrong "you're compliant" would be worse than an honest "we don't have enough information yet."
Is a signator.ai report legal advice?+
No. Every report carries this disclaimer, and it's worth taking seriously: the decision tree encodes a non-lawyer-reviewed reading of Article 50 based on the European Commission's published guidance. It's informational only. For an actual legal determination, talk to a lawyer.