Deepfakes and chatbots: what Article 50 actually requires

Last July the European Commission published its final guidance on Article 50 of the AI Act, and for the first time it spells out, with concrete examples, something every company had been interpreting on its own: what actually counts as a deepfake, when the satire exception applies, and how explicit a chatbot’s disclosure needs to be. At Evicron, an AI and custom software studio in Barcelona with over 200 projects delivered across 12 sectors since 2019, we already covered back in July that Article 50’s transparency duties kicked in on 2 August without waiting for the rest of the Regulation. What was missing was the fine print: how to tell, in practice, what triggers a disclosure and what doesn’t.

What we already knew: Article 50 doesn’t wait

Regulation (EU) 2024/1689 has required, since 2 August 2026, disclosing when a chatbot is AI, technically marking synthetic content, and flagging deepfakes and AI-generated text on matters of public interest. The enforcement machinery — fines up to €15 million or 3% of global turnover — is live, and in Spain it’s policed by AESIA, working alongside the data protection authority and sector regulators. We covered that timeline in detail in our earlier piece; here we focus on what’s new.

What is (and isn’t) a deepfake, according to the Commission

The guidance, open for public consultation until 3 June and published in final form on 20 July 2026, sets a clear bar: to qualify as a deepfake, content must bear a close resemblance to something real or plausible. Based on how law firms that reviewed the text read it, an AI-generated image of a person flying unaided, a dragon, or a talking animal does not meet that bar, because no reasonable person would mistake it for reality. A video putting words a politician never said in their mouth, or audio cloning a real broadcaster’s voice, clearly does.

The test doesn’t turn on whether the publisher meant to deceive anyone — what matters is whether a reasonable audience, which may include minors or older people, could take it for real. That reframes the risk calculation for any business running AI-generated creative: intent is irrelevant, effect on the viewer is everything.

The satire exception isn’t a blank check

Article 50 exempts content that is “evidently” artistic or satirical. The guidance narrows that exception considerably: the satirical nature has to be clearly recognisable from the format, the context and what the audience expects. A political parody that reads as parody at a glance can rely on the exception; an AI-generated ad campaign cannot, even with a humorous tone, because its dominant purpose is commercial. And even where the exception does apply, the disclosure duty doesn’t vanish — the AI origin still has to be revealed, just in a way that doesn’t spoil the piece (a discreet watermark instead of a full-screen banner, for instance).

For any business producing AI-made brand video or imagery — increasingly common in social campaigns — the takeaway is simple: a lighthearted tone doesn’t get you out of the disclosure.

Chatbots: “obviously AI” is no longer an easy excuse

Article 50 also exempts cases where it’s “obvious” to a reasonable person that they’re talking to an AI. The guidance reads that exception narrowly: a chatbot with a human name, a profile photo and a conversational tone doesn’t meet the “obvious” bar just by sitting on a website. If your support assistant doesn’t explicitly state, in its first message, that it’s an AI, you’re probably non-compliant — guidance or no guidance.

The Code of Practice: voluntary, but with a roadmap

On 10 June 2026 the Commission published a Code of Practice on marking and labelling AI-generated content. Signing up is voluntary — it doesn’t replace the legal duty under Article 50 — but it lays out the technical approach the Commission considers reasonable: a layered combination of cryptographically signed metadata and imperceptible watermarks, designed so the AI origin can still be verified even if content is cropped or shared out of context. Joining the Code isn’t proof of compliance on its own, but it’s the reference AESIA will most likely use when judging whether a company acted in good faith.

What to do this week

Three concrete steps, no lawyer required yet:

1. Check your chatbots: if they don’t explicitly state they’re AI on first contact, fix it now — the “obvious” exception leaves little room.
2. Audit AI-generated marketing content: anything photorealistic or impersonating a real person needs a disclosure, satirical or not.
3. Ask your generative AI vendors to document how they technically mark their outputs, ideally following the Code of Practice’s layered approach.

We’re not a law firm and this isn’t legal advice, but it is the ground we work on daily: through our AI consulting for businesses we audit which of your systems fall under Article 50 and how to implement disclosures without breaking your brand experience. If you already know what AESIA is but aren’t sure whether your chatbot or campaigns comply, get in touch — the first discovery session is free and we reply within 24 hours.