Article 50 has applied since 2 August 2026. Here is what it actually requires, who has to do what, where the EU “AI” label comes in, and the exemptions people get wrong. Updated 28 September 2026 with the Commission’s final guidelines and Code of Practice.
Need it done for you? Our Article 50 Transparency Pack writes your disclosures, labelling rules and compliance record in 48 hours.
| Who | Duty | Article |
|---|---|---|
| Provider (builds the AI system) | Design systems that interact with people so they are told they are dealing with AI, unless obvious from context | 50(1) |
| Provider | Mark AI-generated audio, image, video and text in a machine-readable, detectable way (where technically feasible) | 50(2) |
| Deployer (uses AI professionally) | Inform people exposed to emotion recognition or biometric categorisation | 50(3) |
| Deployer | Disclose deepfakes; disclose AI-generated text published to inform the public on matters of public interest | 50(4) |
All of it must be clear and distinguishable, given at the latest at the first interaction or exposure, and accessible (Art. 50(5)). A line in the terms of service or the footer does not meet that.
A deepfake, in the Act, is AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful. That covers far more than political fakes: synthetic talent in an ad, a digital double of a real model, AI backgrounds of a real location, a cloned voice, lip-sync dubbing.
Deployers must disclose AI-generated or manipulated text published to inform the public on matters of public interest (news, politics, public health, public safety). The exception: the text went through human review or editorial control and a natural or legal person holds editorial responsibility. Ordinary brand copy is usually outside “public interest” in the first place. Where it isn’t, keep a dated record of who reviewed what; that record is your exemption.
In July 2026 the Commission published its final guidelines on Article 50 and confirmed the Code of Practice on transparency of AI-generated content. For deployers, the Code provides a common EU “AI” label, placed at first exposure and kept with the content as it is distributed. It also expects internal processes, staff awareness, and a way to fix missing or wrong labels. For providers, it sets out layered marking: metadata, watermarking and provenance.
| Format | Practical placement |
|---|---|
| Film, TV spot, online video | On screen from the first frame where AI content appears, plus the description where the platform allows |
| Short social video | On screen in the opening seconds, the platform’s AI-content setting, and the caption |
| Stills, display, out-of-home | On the image itself, legible at the viewing distance |
| Audio and voice | Spoken disclosure at the start, plus text in the description |
| Carousels and multi-asset | On every asset that contains AI content |
Want this as a one-page rule sheet for your team? Get the free Article 50 labelling checklist.
US states are moving too, unevenly. Texas’s TRAIGA (in force since January 1, 2026) requires disclosure of AI interaction in some contexts; California and Colorado have their own laws with different scopes and dates. Client contracts increasingly ask for AI-use disclosure regardless of law. The common thread: a clear record of which AI tool was used, on what, how it was labelled and who approved it. For the client side of this, see do you have to tell clients you used AI?
Platform by platform: how to label AI content on Instagram, TikTok and YouTube.
The Article 50 Transparency Pack writes your labelling policy (or, for AI startups, your interaction notices), disclosure wording, review procedure, staff guide and compliance record, built from the Commission’s final guidelines and the Code of Practice. Fixed price, $399.
See the pack →