A label only machines can read

Three weeks from now, on 2 December, a transition period ends. Generative systems placed on the market before 2 August will have to mark what they produce: images, audio, video, text. The mark must be machine-readable and must make it possible to detect the content as artificially generated or manipulated. This is Article 50 of the AI Act. The Commission has already published the code of practice and the guidelines. The rules are in place.

I read the code with a single question: who is supposed to understand?

Two layers, two audiences

The system works on two levels. The first concerns whoever provides the generative system: technical signals must be embedded in the content. The code asks for at least two techniques combined, for example signed metadata and an imperceptible watermark, plus a way to verify them. The second level concerns whoever uses the system to publish: they must disclose deepfakes and generated text that informs the public on matters of public interest.

The first layer speaks to machines. The second speaks to people. And the second is the fragile one.

For people, the code proposes an EU icon in three variants: fully generated content, modified content, and a basic icon with a second layer of information. It is optional.

I understand why. A rigid obligation on a symbol, in a market that changes every month, may age badly. But the consequence is clear: we have a precise standard for software to talk to software, and a recommendation for people to understand.

Marking is not explaining

I have seen this film before. It is called consent. For twenty years we treated privacy notices as if publishing them meant they were understood. The law presumed comprehension. It never measured it.

The risk is repeating the pattern. The content is marked, so the person is informed. The metadata is there, so transparency is done. But transparency is not a property of the file. It is an event in the mind of the viewer. It happens or it does not.

A label nobody notices is a box ticked. An icon nobody can interpret is decoration. And even an icon that is seen and understood may not be enough if it arrives after the image has already done its work: after the share, after the outrage, after the decision.

That is why I keep repeating one line in my work: comprehension can be measured, the mind cannot be read. You do not need to get inside anyone's head. You need to check, with method, whether a signal is noticed, whether it is understood, whether it changes how a person judges what is in front of them. That is what we do at Neuro Legal Cortex with the reading of legal texts. It can be done with an icon too.

What I would do on 3 December

If I ran a company that publishes content, I would not stop at technical compliance. I would do three things.

First: adopt the EU icon, even if optional. One shared signal is worth more than ten proprietary ones. Fewer choices for the viewer, more freedom of judgment.

Second: place it where it matters, when it matters. Not at the bottom of the page, not in an info menu. Next to the content, before the share button.

Third: test it with real people. A small group of users, a few minutes, one question: do you think this image is real? If the answer does not change with or without the icon, the icon is not doing its job.

Whoever designs the environment in which we choose should carry the burden of proving that it works. Not because a rule says so, at least not yet. For a simpler reason: a label is only useful if someone understands it.

On 2 December, software will be able to recognise generated content. People, not necessarily.