The transparency obligations set out in the European regulation on artificial intelligence came into force this month. They require, among other things, that a person interacting with an AI system be informed of this, and that generated or manipulated content be identifiable as such. This is the first continental framework of its kind, and it directly touches on a question we addressed on Friday.
In our Friday article, we asked whether one has the right to know one is talking to a machine. The European legislator has answered, and it is worth looking closely at what it decided.
What the rules require
Three main obligations, and they need to be distinguished because they do not concern the same actors.
Information about the interaction. A system designed to interact directly with people must be designed so that those people are informed that they are addressing an AI, except where this is obvious given the circumstances.
Content labelling. Providers of systems that generate synthetic content must ensure that these outputs are identifiable as artificially generated or manipulated, in a machine-readable format.
Disclosure for certain content. Those who deploy systems producing content that imitates real people, or texts published to inform the public on matters of general interest, must disclose this.
This last category is notable: it concerns the media, and we apply it like everyone else.
This caveat is essential and often overlooked. It means the obligation does not apply when the context makes the nature of the system manifest. No one needs an assistant clearly presented as such to remind them at every message that it is a machine. That is a good thing: an obligation that produced notices everywhere would become invisible through saturation, exactly like consent banners. The legislator targeted deception, not formality.
What this fixes, and what it does not fix
This entry into force has a real effect: clear-cut cases of deception become sanctionable. A service that deliberately lets someone believe they are talking to a human, or content imitating a real person without disclosure, now falls within scope.
But two difficulties we have documented remain entirely.
Technical verification. Machine-readable labelling runs into what we described with the watermark bypassed in four hours. An obligation does not create a technique that does not exist. For text, the signal remains statistical and fragile.
Scope. An open model downloaded and run locally escapes the framework in practice, as we noted regarding the California law. Someone who wants to deceive deliberately will not use a service that tracks them.
The point we find most sound
The European text makes a choice we had defended on Friday without knowing it was already written in: it targets the relationship rather than the tool.
The obligation concerns leading someone to believe there is a human presence, not the use of assistance. No one has to declare that a text was reviewed by a machine. Everyone must declare that an interlocutor is not someone.
That is the criterion that seemed most solid to us, and it avoids the trap of the blanket notice that no longer informs anyone.
What to take away
This framework will not solve the detection problem, and that is not its purpose. What it does is shift the burden: rather than technically tracking every piece of content, it makes the deployer responsible.
That is probably the right approach, and it is the one we suggested: we stop chasing the machine and return to the one who uses it. A rule that cannot be technically verified can still be effective if it exposes the one who breaks it.
The question of effective enforcement remains. A text only exists through its application, and Europe has a mixed history on this point. We will know in a year whether this obligation has changed anything, or whether it has joined the list of rules everyone knows and no one enforces.