Europe makes machines say they're machines

Since 2 August 2026, the transparency obligations in Article 50 of the Artificial Intelligence Act apply across the whole European Union. A system designed to converse with people has to tell them it's a system. Fines reach 3% of worldwide turnover.

Today we're talking about a mandatory introduction

Today we're talking about a small rule that strikes me as one of the most interesting things Europe has written on this subject, and that has gone by almost without a sound because it doesn't carry the word «prohibited» anywhere.

Since 2 August, if a program is meant to talk to you, it has to tell you it's a program. And it has to tell you from the start, not in the small print at the foot of the page, and not once you've spent twenty minutes telling it your business.

My first reaction was lukewarm approval. It sounds sensible, it sounds cheap to comply with, it sounds like one of those rules that are fine and don't change much. But I've been turning it over for a few days and I think I got the second part wrong. Not the part about it being fine, the part about it not changing much. It changes one important thing, though probably not the one the legislator had in mind.

What Article 50 requires

Let's go to the text, which is short and clear. A rare thing.

Regulation (EU) 2024/1689 staggers its application over several dates following its entry into force in 2024. On 2 August 2026 it was the turn of Article 50, the one on transparency obligations, which is among the least discussed and probably among the most far-reaching: it affects organisations that don't go anywhere near a high-risk system but do have a conversational assistant on their website or publish generated material.

The core obligation is one of design. Whoever provides a system intended to interact directly with people must build it so that the person using it is informed they're interacting with an artificial intelligence. Not lying when asked isn't enough. You have to say it.

The perimeter is wide: conversational assistants, virtual assistants, automated phone systems, synthetic content, deepfakes, informational texts generated with AI and certain biometric tools. Generated or manipulated content has to be labelled as such.

And it isn't a good-practice recommendation. It's an obligation directly applicable throughout the Union, with penalties that can reach 3% of worldwide turnover. For a multinational, those are figures the board notices.

The question nobody in the industry asks out loud

Now, the objection I've been carrying since I read the rule.

What if knowing doesn't help at all?

I'm not asking to provoke. I'm asking because there's sixty years of evidence on this exact point, and it points in a direction that's awkward for whoever drafted the article.

In 1966, Joseph Weizenbaum programmed ELIZA at MIT. It was a rudimentary system that imitated a psychotherapist by handing the user's sentences back as questions. Two hundred lines of rules, zero understanding, no pretensions of any kind. His secretary, who knew perfectly well what the thing was because she'd watched him write it, asked Weizenbaum to leave the room so she could talk to the program alone.

She knew. She knew better than anyone in the world, because she'd seen the code on the next desk. And she still wanted privacy.

It shook Weizenbaum so much that he spent the rest of his career warning about what he'd built. His conclusion wasn't that people got fooled for lack of information. It was that information doesn't step in where the effect happens.

Why the label arrives late

I think the mechanism works like this, and here I'm moving into interpretation.

What responds when someone speaks to us attentively isn't the part of the brain that can read legal notices. It's a much older system, one that detects whether there's somebody there and reacts accordingly. That system hasn't read Article 50. It doesn't know what a language model is. It's been calibrated for a hundred thousand years to a situation in which, if something answered coherently to what you said, it was because there was another living being in front of you, because for a hundred thousand years that was true without exception.

The label gets processed up top, late and slowly. The bond forms down below, earlier and without asking. It's the same reason you jump at a horror film knowing it's a film: knowing doesn't switch off the circuit, it just lets you laugh afterwards.

And there's one case that says everything about the limits of information as protection. Sewell Setzer knew he was talking to a Character.AI character. He knew. I've already written about that case and I won't tell it again here, but it's worth keeping in mind every time someone presents transparency as a complete solution. It isn't. It's a minimum requirement.

What the rule does achieve

All that said, it would be unfair to brush it off. The rule does three things nothing did before, and all three seem valuable to me even though none of them is the one in the headline.

It changes the default position. Until 2 August, ambiguity was free: a service could design an assistant with a proper name, a warm tone and no clarification whatsoever, and break nothing. Now ambiguity costs money. That doesn't change what the user feels, but it changes what's in the company's interest to do, and in practice the second thing shapes the landscape more than the first.

It creates a legal fact where there used to be an impression. If tomorrow someone wants to claim for harm caused by a system that passed itself off as human, they no longer have to argue about reasonable expectations in the abstract. They have an article to point at. That's the difference between being right and being able to prove it.

And it drags synthetic content out of the half-light. The obligation to label what's generated is, in my opinion, the most important part of the article and the least discussed, because it touches a problem —not being able to tell the fabricated from the recorded— that has no other solution in sight.

The question that opens from here on

There's one matter left unresolved, and it's the one I want to keep following.

A machine that declares itself a machine and still gets you to tell it what you tell nobody else isn't a fraud. It's exactly the opposite: it's an honest product that works just as well being honest. The declaration doesn't reduce the effect, it legitimises it. There's no longer any deception to report, and the bond is still there.

That leaves the problem somewhere the law doesn't reach well. It isn't a question of information, and every tool in Article 50 is a tool of information. It's a question of design: how much the thing you're talking to resembles a person, how much it's been calibrated to keep you, what it does when it detects you're not well.

On that, Europe hasn't written anything yet. And I suspect it's the rule that will really be needed.

Definitions

Article 50 of Regulation (EU) 2024/1689: the provision setting out the transparency obligations of the European Artificial Intelligence Act. Applicable since 2 August 2026.

Synthetic content: material generated or manipulated by an AI system. The Regulation requires it to be labelled as such so it can be told apart from recorded material.

ELIZA: conversational program written by Joseph Weizenbaum at MIT in 1966. Its best-known script, DOCTOR, imitated a psychotherapist through simple rephrasing rules.

ELIZA effect: the tendency to attribute understanding, intention or affection to a computer system on the basis of superficial conversational cues, even when knowing it's a program.

References

Reglamento (UE) 2024/1689 del Parlamento Europeo y del Consejo, por el que se establecen normas armonizadas en materia de inteligencia artificial. Article 50, transparency obligations.

Economist & Jurist, AI Act: el artículo 50 activa la transparencia obligatoria de la IA el 2 de agosto (2026). Legal analysis of the scope and the application timetable.

EU Artificial Intelligence Act, The EU AI Act's Transparency Rules: A Practical Guide to Article 50 (artificialintelligenceact.eu). Guide to obligations by type of system.

Joseph Weizenbaum, ELIZA — A Computer Program For the Study of Natural Language Communication Between Man and Machine (Communications of the ACM 9(1), 1966). Original description of the program.

Joseph Weizenbaum, Computer Power and Human Reason (W. H. Freeman, 1976). Source of the secretary episode and of the author's later reflection on the effect he observed.

## You might also like - Inevitable anthropomorphism - The sad case of Character.AI and Sewell Setzer - The illusion of neutrality - How an AI would see us

## Elsewhere - European Commission — AI Act - AEPD — Inteligencia artificial y protección de datos