The students’ involvement began through a study project on AI in debt counselling, during which they came into contact with Femke de Lange, a lecturer in Higher Professional Law and coordinator of the minor Human-Centred Digital Government. Within that minor, students work on advisory reports examining how the government can deploy digital tools responsibly. Max and Stijn initially gave a guest lecture for the minor, but their role expanded and they ultimately made a substantial contribution to the report itself.
The Council of State holds a central position within the Dutch government: the body advises the government and parliament on legislation and governance, and also serves as the highest general administrative court in the country. Large volumes of documents and files are processed daily. According to Max, the workload at the Council of State is high, with a significant portion of the work consisting of searching through legal documents. The central question in the research was therefore not only whether AI can assist with this, but above all how it can do so safely and responsibly.
A prominent theme in the research was the dependency on large international technology companies. Many government organisations work with systems from American companies such as Microsoft. Stijn points out that when working with sensitive data, careful thought must be given to where that data ends up and who has access to it. A concrete risk is the American CLOUD Act, which can in certain cases grant American authorities access to data held by American companies, even when that data concerns Dutch information stored on servers located in Europe. Max notes that this theoretically means data from Dutch government organisations could be accessible to foreign authorities, which could have serious consequences for an organisation like the Council of State.
A key conclusion of the advisory report is that organisations such as the Council of State should keep technology in-house as much as possible. The students highlight GPT-NL, a Dutch language model trained largely on recognised Dutch data and running on Dutch servers, as a concrete alternative. Connecting such a model to internal systems creates a closed infrastructure in which data is not sent to external parties. At the same time, Max and Stijn acknowledge that switching to such alternatives is not straightforward in practice. Many organisations have been embedded in the ecosystems of large technology companies for years, and even when an alternative functions well technically, that does not automatically make the transition simple.
