AI chats and Dutch FOI law: what public bodies keep

On a Tuesday afternoon, a policy officer at a Dutch municipality asks Copilot for three versions of a council proposal on the new parking scheme. She asks which objections residents are likely to raise, and which wording will meet the least resistance. She picks the second version, polishes it and sends it to her team leader. The chat stays in the side panel.

Two months later, a journalist asks for every document on how the parking scheme came about. Does that conversation with Copilot count? The short answer: probably yes. Then you want to know where it is stored, who can find it and whether it still exists. That depends on arrangements you can make today.

Why a conversation with AI can be a document

The Dutch Open Government Act (Wet open overheid, Woo) does not mention AI anywhere. It does not need to. Article 2.1 of the Woo defines a document as including any “set of recorded data” that by its nature relates to the public task. A chat about a council proposal is recorded and concerns the public task. Whether it was a chatbot or a colleague is not part of the definition.

For text messages, this debate has already been settled. In 2019 the Council of State, the highest Dutch administrative court, ruled that work-related SMS and WhatsApp messages are documents, even on a private phone. The Dutch central government’s policy on archiving chat messages puts it briefly: the form of the information does not matter to the law. According to that policy, messages about carrying out public tasks are also records under the Public Records Act (Archiefwet).

Public bodies are already applying this to AI. The Midden-Holland environmental agency states in its policy on AI use that documents made with AI fall under the Public Records Act and the Woo. In Rotterdam, a council group submitted written questions in September 2025 asking whether Copilot and ChatGPT were “the new text message debacle”. The question is on the table.

What happens when someone asks for it?

The United Kingdom has already shown it can happen. In February 2025, someone used the UK Freedom of Information Act to request the ChatGPT history of minister Peter Kyle. A month later, his department published the prompts and responses. Among other things, he had asked which podcasts he should appear on. Different law, different country, but an ordinary request.

In the Netherlands, a public body has four weeks to decide, plus at most two weeks’ extension. More important is Article 4.1a: once a request comes in, the body must make sure the requested documents are preserved. Cleaning up a chat history after the request has arrived is therefore not an option. If you clear your chats every Friday, you need to stop straight away.

The usual exemptions still apply. Personal policy views in documents for internal deliberation are not disclosed (Article 5.2), and personal data is redacted. But an exemption is not the same as “that doesn’t count”. The conversation first has to be found and assessed, just like an email or a draft memo.

Where is that conversation actually stored?

That depends on the tool. With Microsoft 365 Copilot, prompts and responses are kept, according to Microsoft, in a hidden folder in the mailbox of the user. An administrator can search them with eDiscovery and apply a retention policy. With such a policy in place, a chat the employee deletes is not gone straight away.

A free ChatGPT account on a personal laptop is a different story. The conversation sits with an American provider, outside any archive of the organisation. Anyone who needs it has to ask the employee for an export. Text messages on private phones were already found to be documents. It seems reasonable to expect the same for such a chat.

Why does that matter? Because an organisation that does not know where its documents are cannot fully answer a freedom of information request. That is exactly what went wrong with text messages. We wrote earlier about which data in which AI tool you may use. This question is the other side: what has to come back out of that tool.

What to arrange now

The Dutch government’s guidance on responsible use of generative AI is not binding, but it is clear. Make agreements on whether prompt history is stored, “for transparency and traceability”. And if you use AI to create a government product, keep the prompts and the output available for those involved and for review. In practice, that comes down to four things:

1. Choose a tool whose archive you control. A business environment with a retention policy is easier to account for than loose accounts. That keeps chats from disappearing out of sight.
2. Decide what goes into the file. Not every conversation is worth archiving. A chat that contributed to a decision, like the versions of that council proposal, belongs in the case file.
3. Include AI in your retention schedule and FOI procedure. Whoever handles a request needs to know that Copilot chats exist and where to find them.
4. Tell your staff. Most civil servants do not realise their chat is a document. That belongs in training for policy makers and the public sector, alongside prompting and privacy.

It does not have to be a major operation. Often it is a few lines in your AI policy, a setting for the administrator and one paragraph in the instructions for handling requests. Harder is deciding which tools you allow, and what to do about the free accounts already in use.

In short: a conversation with AI about your work is, for Dutch freedom of information and archives law, a document like any other. The difference lies in how easy it is to find, and you settle that with the tool and a few agreements. If you want to set this down properly, our AI policy consulting helps with the agreements, and our AI training for municipalities with day-to-day use.

Frequently asked questions

Does every conversation with ChatGPT or Copilot fall under the Woo?

No. It concerns conversations that by their nature relate to the public task. Asking for a weekend recipe falls outside it. A conversation in which you prepare a decision, memo or letter will usually fall within it.

May I delete my chat history?

That depends on your organisation’s agreements and its retention schedule. What must be archived cannot simply be thrown away. And once a request on the subject has come in, the public body must preserve the documents.

Should I state that a text was made with AI?

The Dutch government guidance recommends it: name the system as a source of information when you use AI to create a government product, and keep the prompts and output available for review. Make clear in your own policy when you do this.

Does this also apply to a free account on my own phone?

For text messages, the Council of State ruled that it makes no difference whether they are on a work or a private phone. The same reasoning seems likely for an AI chat about your work. The safest choice is to use only the tools your organisation has approved for work.