What BesluitBron is under the European AI Regulation, Regulation (EU) 2024/1689. This page explains which parts of the product constitute an AI system, which parts do not, and what this means for those who use it. It is written for readers rather than regulators, as readers are usually civil servants who need to be able to explain what they are using.
The transparency rules set out in the Regulation have been in force since 2 August 2026. The obligation to ensure that people working with AI understand what they are using has been in force since 2 February 2025.
Three parts, three answers
BesluitBron consists of three elements that are often lumped together. Under this regulation, there are actually three of them.
The connectors are not an AI system. They search public sources and return whatever information is found there. There is no model involved that infers or predicts anything: a search query with filters is not AI, however clever the result may appear. See Connector.
The question-and-answer service is indeed an AI system. It uses a language model to answer questions. BesluitBron builds this service, offers it under its own name and operates it, and is therefore both the provider and the data controller. The obligations regarding transparency therefore lie with BesluitBron. See Vraagbaak.
The assistant that someone connects themselves is not provided by BesluitBron. Whoever connects Claude or any other assistant to the connector is responsible for that system themselves. BesluitBron simply provides the tools for this – and nothing more.
That distinction is not just a play on words. It determines who is required to report what, and it prevents BesluitBron from claiming a role that it does not have.
What it is intended for, and what it is not intended for
The regulation makes the stringency of the rules dependent on the purpose a product claims to serve. That is why that purpose is set out in the code, in one place, and appears wherever it is to be read: in the terms of use, in the response from list_capabilities and in the instructions an affiliated assistant receives upon joining.
The aim is to make public decision-making documents searchable and citable, so that anyone can go to the source and read it for themselves.
Three types of use are expressly excluded from this:
- use by a court, or by any person acting on its behalf in any capacity whatsoever, and the same in the context of alternative dispute resolution;
- used to influence the outcome of an election or a referendum, or people’s voting behaviour;
- used to determine whether someone is entitled to a public service or a benefit.
The first exclusion has deliberately been defined more broadly than the regulation requires. This immediately brings into play the other side of the coin, which is just as strict: a party to proceedings, their lawyer and their adviser are, in fact, included. The exclusion concerns the judge themselves and those who work for the judge, not those who appear before the judge. This tool is designed to assist with reading a document and is never intended to aid in deciding a case.
What a reader of the advice column sees
There are two notices on the form, in the reply and in the email. They serve different purposes and one does not replace the other:
- The answer has been generated by AI and may contain errors;
- This is a test facility, and the response is not intended for reuse.
The first is the obligation set out in Article 50. The second is a statement regarding the service. “A test, not for reuse” does not imply that a machine wrote the text and therefore does not remove that obligation.
In addition, a reply page bears the header ‘AI-Generated: true’, and the reply email has the same header. This is intended for machines.
What isn’t finished yet, and why it’s here
The regulation requires that generated text be marked in a machine-readable format. The associated code of practice requires two layers for text on a web page: metadata with a digital signature, and an invisible watermark. BesluitBron provides one of these – the watermark that comes with the template. The layer containing the signature is not generated.
That is a choice, not a mistake, and it is included here because any reader who visits this page is entitled to the full answer, not just the part that puts us in a favourable light.
Those who operate it must know where things go wrong
The regulation also requires that people working with these systems understand where the limits lie. For BesluitBron, this has been set out for two groups: those who operate and are responsible for the knowledge base, and those who build this product using an AI assistant. These two groups encounter different issues. The former must know where the answers are unreliable, whilst the latter must know where the model makes mistakes regarding its own output.
Three things everyone who runs the helpdesk needs to know:
- A well-formulated answer is not proof;
- "Nothing found" means: nothing was found in what was searched, within the window that was searched. See Zoekprecisie;
- The sources are incomplete. Three-quarters of the rulings have never been published, and the data feed is interrupted on several source platforms. See Beta en Volledigheid.
Read more
The full analysis, including the considerations and outstanding issues, constitutes project documentation rather than product documentation. Anyone wishing to obtain a copy may request it via the address given at Contact.
Available from version 1.3.41. The section on the intended purpose is available from version 1.3.50.