GPT-NL Brings Generative AI to Dutch Municipalities

Language-based AI is increasingly used in work, education, and public services, but who develops and governs it directly affects cybersecurity, data protection, and societal trust. GPT‑NL takes a different approach, prioritising oversight, transparency, and public interest.

In 2025, TNO, NFI, and SURF trained GPT-NL on the National Supercomputer Snellius using exclusively lawfully sourced data, positioning it as a sovereign and responsible alternative to existing large language models.The model has even received the Dutch Privacy Award for its focus on transparency, responsible data use, and compliance with intellectual property rights.

GPT‑NL has been installed on ICTU’s own AI infrastructure to test the virtual municipal assistant Gem. ICTU, 27 municipalities, and TNO are the first testers, aiming to further strengthen the Netherlands’ position as an “AI country.” Several municipalities, including Utrecht, Rotterdam, and Eindhoven, already use Gem to provide answers about municipal services such as civil affairs.

From development to real-world application

The Gem community has six years of practical experience, serving around 2.8 million residents. However, Gem has not yet integrated generative AI, limiting functionality and increasing maintenance costs. By introducing GPT‑NL, the partners aim to enhance Gem’s capabilities, expand service offerings, and reduce costs. Testing with Gem represents the transition of GPT‑NL from early development into practical application, helping refine the model and explore its potential to improve public services. ICTU expects to share results after the summer of 2026.

Transparency, security, and public trust

Openness is a core principle of GPT‑NL. The source code is published as open source, with controlled licensing of model weights and clear documentation on dataset composition and design choices. Personal data is anonymised, confidential and harmful content is excluded, and intellectual property rights are protected. Developed entirely within the Netherlands and Europe, GPT‑NL ensures control over the model, data, and deployment, reducing reliance on overseas providers while aligning with European legislation and societal values.

Trustworthy: protecting users and citizens

The model is developed entirely from scratch. This approach prevents the transfer of unclear provenance, copyright risks, or potential personal data from existing models. To ensure a secure and reliable foundation, the data collection process adheres to strict criteria:

• Protection of intellectual property

• Removal and anonymisation of personal data before training

• Exclusion of information subject to confidentiality requirements

• Exclusion of harmful content

• Avoidance of duplication within the dataset

Responsible and effcient AI

GPT‑NL is built on lawful and ethically sourced data, with oversight by a dedicated Content Board that ensures fair agreements with creators. Energy efficiency and responsible use of computational resources are prioritised, guided by scientific research. Funded publicly by the RVO, the Ministry of Economic Affairs and Climate Policy, and for the Gem tests, the Ministry of the Interior and Kingdom Relations (BZK), GPT‑NL demonstrates that powerful AI can coexist with public values, security, and accountability.

By integrating GPT-NL with Gem, the Netherlands is taking a concrete step towards digital autonomy, secure municipal services, and responsible AI innovation. It demonstrates that generative AI can strengthen public services while respecting societal and legal frameworks.

Would you like to learn more about the current state of GPT-NL? Read the progress report here: tno.nl

Photocredits: Istock/Luechai