you are talking to a machine

The label is an intervention, not a disclaimer.

On August 2, 2026, a new sentence entered the regulatory lexicon of every chatbot operating in the European Union: You are talking to an AI.

Article 50 of the EU AI Act requires it. Any company deploying a system that interacts with people through conversation must disclose, at the start of the interaction, that the other side isn’t human. Plain language. Accessible format. No burying it in terms of service. The penalty for skipping this step: up to €15 million or 3% of global annual turnover, whichever is higher.

The requirement is narrow compared to the rest of the Act. The Digital Omnibus package pushed high-risk AI system compliance out to December 2027. But these transparency obligations — chatbot disclosure, synthetic content labeling — went live on schedule. The AI Office now has supervisory and fining authority over general-purpose AI providers to enforce them.

So here we are. Every chatbot in the EU now opens with some version of a confession.

The obvious question is whether companies will comply. The real question is what compliance accomplishes.

There is a philosophical tradition that takes this seriously. J.L. Austin called them performative utterances — sentences that don’t describe the world but change it. “I promise” doesn’t report a promise. It creates one. “I now pronounce you married” doesn’t observe a marriage. It performs it. The words don’t point at something. The words do something.

“You are talking to an AI” looks like a description. A piece of information delivered before the conversation begins. But watch what happens when you say it. The interaction changes shape. The user recalibrates expectations, adjusts their trust model, decides what questions are worth asking. The disclosure isn’t metadata about the conversation. It’s the first move in the conversation.

This is the thing the regulation gets right, even if it doesn’t know it. The label is an intervention. It restructures the relationship before any content is exchanged. You don’t read the words and then decide what to do with them. The words reorganize the space you’re standing in.

But interventions have to intervene on something. And here the logic starts to thin.

The disclosure protects people who already suspect they might be talking to a machine. It gives them confirmation. It resolves ambiguity. For this group — and it’s a large group, growing larger every month — the label is a courtesy, not a revelation.

The people the regulation ought to be protecting are the ones who would never have guessed. The elderly person who thinks the customer service agent has been patient and kind. The teenager who doesn’t distinguish between a chatbot therapist and a human one. The voter who reads a generated news summary and takes it as reported fact.

These people are not reading disclosure banners.

The regulation assumes a model of harm that runs through knowledge. If people knew, they’d behave differently. If they behaved differently, they’d be protected. The intervention is information, and information is sufficient.

That assumption is testable. The early evidence is mixed. Some research suggests that labeling a conversation as AI-generated reduces trust in the content. Other work suggests it makes no measurable difference in how people engage, what they believe, or whether they act on what they’re told. The label changes the frame. It doesn’t always change the picture inside it.

There is a pattern here that anyone who works in compliance recognizes. The regulation targets the interaction that is easiest to regulate — the one where the deployer is identifiable, the interface is controlled, the label can be inserted at a known point. The chatbot on a corporate website. The customer service flow. The clearly bounded product.

It does not reach the places where the confusion is most dangerous. A voice clone in a phone call doesn’t pause to introduce itself. A generated image circulating on social media doesn’t carry its provenance with it by default. Article 50 requires synthetic content labeling too — but enforcement requires detection, and detection requires infrastructure that does not yet exist at scale.

The regulation draws a line. On one side: interactions that are structured enough to label. On the other: everything else. The line isn’t wrong. It’s incomplete.

None of this means the requirement is pointless. A norm has to start somewhere.

There was a time when ingredient labels on food seemed performative in exactly this way — a gesture toward transparency that most consumers ignored. Over decades, those labels built a literacy. People learned to read them. Institutions learned to enforce them. The labels didn’t work because individuals acted on them in the moment. They worked because they established a standard that other systems could build on.

Maybe that’s what Article 50 is doing. Not protecting anyone today. Establishing a floor. Making disclosure the default so that its absence becomes meaningful. The value of the label isn’t in what it tells you. It’s in what it costs to remove.

The AI Office has fining powers now. The question isn’t whether the first fines will come — they will. The question is whether the fines will be for the interactions that matter, or for the ones that are easy to catch.

A chatbot tells you it is a chatbot. You nod. You type your question anyway.

The disclosure changed nothing about the words on the screen. It changed everything about what you’re willing to hold them responsible for.