Field test: do Dutch chatbots tell you they are AI? Ten organisations tested against Article 50

On 13 July 2026 we asked the same everyday service question to the chatbots of ten major Dutch organisations. Three explicitly tell you that you are talking to AI: NS (Dutch Railways), Ziggo and health insurer CZ. Two call themselves a chatbot without using the word AI. Three choose softer labels such as digital assistant, virtual assistant or digital service employee. Two had no publicly reachable chat. Since 2 August 2026, Article 50(1) of the EU AI Act requires that people who interact with an AI system are informed of that fact.

We covered the deadline itself earlier: it stands and has not been postponed by the Digital Omnibus. This piece is different in kind. It is not an explanation of the rules but a snapshot from practice: how did the chatbots of major Dutch organisations introduce themselves on 13 July 2026, three weeks before the obligation started to apply?

How did we test?

The setup was deliberately simple and repeatable. On 13 July 2026 we opened the public customer service chat of ten major Dutch organisations, as an ordinary visitor, without logging in. We asked the same kind of everyday service question everywhere and recorded the exact opening and disclosure wording, with screenshots as evidence.

Two ground rules. Organisations that do this well are named: that is a compliment and a usable example. Patterns that raise questions are described anonymously, because Article 50(1) contains an exception for situations where the AI nature is obvious from the context. According to the Commission guidelines that exception is narrow: there must be almost no doubt (guidelines, point 45), and a generic label such as 'assistant' is not enough on its own (guidelines, point 38). None of the findings below is an established infringement, not least because the test took place before 2 August 2026, the day the obligation started to apply.

What does Article 50(1) actually say?

Providers of AI systems intended to interact directly with people, such as chatbots and voicebots, must design those systems so that the person knows they are communicating with AI. The obligation does not apply where that is obvious from the point of view of a reasonably well-informed, observant and circumspect person, taking into account the circumstances and the context of use.

For the full explanation of the obligation, including the division of roles between provider and deployer, see does your chatbot have to say it is AI? and provider or deployer: who has to arrange what?

What did the ten chatbots tell us?

The three explicit examples deserve to be named, because they show that the solution is not complicated.

NS is the gold standard of this test. Before you type anything, a separate screen titled "Privacy and AI" opens. It states that Jens is an AI chatbot, that the answers are generated by AI based on internal information sources, that AI can make mistakes, and how to reach a human agent. Including a privacy explanation and a link to the privacy statement.

Ziggo takes the shortest route that works. The chat window is called "Ziggo AI Assistent" and the very first line of the conversation reads, in Dutch: "Onze chat gebruikt AI" (our chat uses AI). Four words, no room for misunderstanding.

CZ labels the answer rather than the channel. The Q&A function on its service page places under every generated answer: "This answer was automatically generated with the help of AI. The information may be incomplete. When in doubt, always check the source or contact us." Notably, CZ deliberately keeps its live chat with human agents and says so. Such a visible per-answer label is not itself required by Article 50: a chatbot answer seen only by the person asking is not published text under Article 50(4) (guidelines, point 131). The provider must, however, mark the text in a machine-readable way under Article 50(2), covered in machine-readable marking of AI content.

Is "digital assistant" enough?

Half of the organisations we tested sit in what we call the grey zone. The bot is called a chatbot, digital assistant, virtual assistant or even digital service employee, but the word AI appears nowhere. Is that a problem now that the obligation applies?

The honest answer: for a public customer service chat, reliance on the exception is weak. The guidelines expressly call a generic label such as 'assistant' insufficient (guidelines, point 38) and limit the exception to cases with almost no doubt; for a public system that children, older people or persons with disabilities may use, you cannot rely on it (guidelines, points 44 and 45). The defence is that a reasonably observant user who sees a chat window labelled chatbot or digital assistant understands that no human is answering. That argument weakens as the bot is dressed up more like a human: a first name, a face as an avatar, being referred to as "she", or a job title such as service employee that normally denotes a person. An organisation that deliberately humanises its bot while leaning on the context exception is asking a lot of that same context.

The supervisor is not the only risk either. Fines run through the national supervisory authority, a route we describe in enforcement and fines under Article 50, but the faster risk is reputational: customers who find out after 2 August that the friendly service employee was an AI system will mostly remember that nobody told them.

What should you arrange now?

The good news from this test: the gap between the grey zone and explicit disclosure is small. NS, Ziggo and CZ prove that a single design decision is enough. In practice it comes down to three steps.

What happens next?

This field test will get a sequel. Now that the obligation applies, we will repeat the test at the same ten organisations and publish what has changed. We will also extend the test to government chatbots, where the trust question weighs even heavier.

For the broader picture of all four transparency obligations, start with Article 50 in practice: all obligations at a glance.

Want to determine which obligation applies to your own use case? Walk through the Article 50 decision tree: per role and use case the paragraph, the steps from the guidelines and examples of good notices.