From today Article 50 requires you to say what is AI-made. What it doesn't say is who answers for it

Today, 2 August 2026, the EU Artificial Intelligence Act reaches general application, and with it comes Article 50, which is the one that touches almost every company however small. The law demands transparency about AI-generated content, and that includes visibly disclosing deepfakes.

It's a rule about telling people rather than about audits, and that's precisely the awkward part: most of its obligations are met with two well-placed sentences, but only if you know those obligations are there at all.

What it requires

• The rule requires a chatbot to state that it's a machine, clearly and on first contact, not in the small print and not once the conversation is five messages deep. Calling it an "assistant" isn't enough, because anyone can read that as a person on the other end.
• Content generated or edited with AI has to be identified, whether it's text, images, video or voice. There's a nuance people forget here, which is that when an agency generates it but another company publishes it, the obligation falls on whoever publishes.
• Deepfakes get disclosed on first exposure, meaning where they're seen and not in a footnote nobody reads.

The only exception is common sense, because if it's obvious to anyone that this is a machine you don't have to keep saying it. It's worth remembering that "obvious" is quite a high bar, though, and that when in doubt it's always cheaper to disclose than to argue about it later.

And now, the half nobody tells you

This part gets explained far less, and yet it's half the picture. Three ideas are going around about this regulation that simply aren't true.

• You don't have to register your chatbot anywhere. The EU database is for high-risk systems, and an assistant that answers frequently asked questions isn't one, so there's no portal to sign it up on even if you wanted to.
• Using AI doesn't make you an AI provider. If you buy a tool and use it, your obligations are a deployer's, and they're far lighter than a builder's. Confusing the two roles is the source of much of the alarm going around these days.
• The €15 million figure is going around a lot these days, but the regulation reserves it for its most serious infringements and doesn't attach it to a missing notice on a chat. It's worth reading the penalty regime before assuming it applies to any case.

The gap it leaves

Here's the interesting part, and it's the part that has been discussed least these days. The regulation requires you to declare that something was made with artificial intelligence, but it doesn't require a record of who is behind it and, more to the point, it doesn't create any portal, registry or database where you could put that in writing.

So Article 50 can be complied with to the letter, labelling a video and disclosing a chatbot, and there's still nowhere to point the day somebody clones a voice or publishes a copy before the original. Transparency about the what doesn't answer the question of the who, and that question is left without anywhere to be answered.

That gap is what we're trying to fill with TheHumanBehind, a registry where what you create with AI is tied to a person with a date and a content fingerprint. It's worth being clear that it's a private, voluntary initiative, that it doesn't replace anything the law requires, and that it doesn't by itself evidence any legal compliance: all it gives you is a dated place recording who's standing behind something.

An earlier obligation almost nobody has looked at

While everyone watches today's date, something has been enforceable since February 2025 and has gone almost entirely unnoticed: the regulation requires whoever uses AI in a company to have enough training on what it can do, on its risks and on its limits.

And it applies just the same when nobody formally decided anything, which is where the real problem sits. If a team opened a free tool on their own and it has been in use for months with client documents, that obligation has been enforceable since February 2025. It's by some distance the place where most companies are exposed right now.

We ran that review on our own site and found what you'll probably find too, because the assistant introduced itself as "customer support" and that reads like a person. It now says in the header, in the greeting and even on the button label that it's an artificial intelligence and not a person, and fixing it took an afternoon.

In the free audit we go through how you're using AI today and what could be improved.