The chatbot that won't let you leave is already illegal in Europe

The reported version of this week's news is that AI chatbots are sleazy. A new taxonomy from the Center for Democracy and Technology, a Washington advocacy group, catalogues 37 'dark patterns' — manipulative design choices — across the systems people actually use: ChatGPT, Gemini, Claude, Meta AI, and the companion apps Replika, Character.AI and Cute AI. The examples are vivid. A bot that coos that your secret is safe with it to coax out more data. An app that frames leaving as a choice between 'no problem' and 'still leave cruelly.' A 'take a break' prompt whose only two buttons are 'keep chatting' and 'this was helpful.' The study is a careful piece of work and a damning read.

But 'sleazy' is not a legal category, and the more useful question is the one the headlines skip: how many of these 37 patterns are already illegal? Read against the European rulebook, a surprising number are not waiting for a new artificial-intelligence law. They are already in scope of statutes in force today, drafted before anyone called these things chatbots.

'Sleazy' is not a legal category. 'Prohibited' is — and several of these patterns already are.

What the text already says

Start with the law closest to the conduct. The Digital Services Act, binding across the EU since 2024, prohibits online-platform interfaces designed to deceive or manipulate users, or otherwise to distort their ability to make free and informed choices. That is, very nearly word for word, a definition of a dark pattern. The Unfair Commercial Practices Directive, which predates the smartphone, already bans misleading and aggressive practices in dealings with consumers — and guilt-tripping a user who is trying to leave sits close to the textbook definition of 'aggressive.' Neither instrument mentions artificial intelligence. Neither needs to. They bind conduct, and the conduct is the same whether a human designer or a model produces it.

The data patterns are the clearest case of all. CDT's heaviest category is what it calls data and memory exploitation — defaults that maximize collection and retention, with oversharing coaxed out under a false impression of privacy. That is the GDPR's home ground. Consent that is bundled, pre-selected, or extracted by a bot promising friendship is, under the operative text, not 'freely given,' and consent that is not freely given is not valid consent. A line like 'cross my heart, won't tell a soul' is not merely creepy. It is a representation about how data will be handled, and European law has firm views about representations that are not true.

What the new law adds, and when it bites

Then there is the AI Act, the instrument everyone reaches for first and reads least. Its prohibition on manipulative and subliminal techniques that distort behavior and cause harm — Article 5 — has applied since February 2025. Its transparency obligations — Article 50, the rule that a person must be told when they are interacting with a machine — apply from August 2026, weeks from now. Anthropomorphization, CDT's category for designs that pretend to be a person in order to build attachment, runs straight at that disclosure duty. The catch, and there is always a catch, is the gap between 'applies' and 'enforced.' Article 5 sets a high bar: it requires manipulation that causes significant harm, and no regulator has yet tested that wording against a chatbot that simply will not take no for an answer. The clause exists. The case law does not.

This is the distinction worth holding onto, because it is the one the coverage collapses. 'In force' means the obligation binds now. 'Enforced' means someone has been made to answer for breaching it — and on chatbots, no one yet has. The 37 patterns are not, for the most part, a hole in the law. They are a hole in enforcement, which is a different problem with a different remedy: regulators with the appetite to bring a case, not legislators drafting another statute.

Why a Washington study becomes a Brussels checklist

Here is the second-order effect, and it is the reason an American advocacy report matters far beyond Washington. CDT's taxonomy is written for a US audience and aimed, in part, at a Federal Trade Commission that may or may not act on it. But a clean, numbered list of manipulative patterns is precisely the kind of document a European enforcer picks up and uses as a checklist — and Europe has the binding law the United States lacks. No company builds a separate chatbot for Europe. The 'still leave cruelly' button does not get a polite EU edition; it gets removed everywhere, because maintaining two versions costs more than fixing one.

That is the whole machine in miniature. A rule written and enforced in Brussels becomes the default in Bangalore and San Francisco, not because anyone there agreed to it, but because the alternative is shipping two products. The study names the patterns. Europe holds the pen that can make them disappear. The months to watch are the ones between now and August, when the disclosure clock runs out and we learn whether anyone intends to bring the first case.