The Netherlands’ ambitious homegrown AI model enters the real world

Five organizations have begun feasibility studies with plans to expand and eventually launch commercially later in the year. One of the first pilots involves Gem, a virtual assistant already used by nearly thirty Dutch municipalities. The feasibility study is examining whether GPT-NL can improve the quality of responses citizens receive when asking Gem questions.

Another pilot focuses on a government writing assistant designed to help civil servants draft clearer letters. That may sound less glamorous than image generation or AI video tools, but it touches on a very real issue in public administration. Official communication around benefits, debt, and social services is often dense enough to confuse the people receiving it. GPT-NL is being tested to see whether it can make those interactions more understandable.

The Netherlands Forensic Institute is using GPT-NL for its work, fine-tuning it on forensic datasets to improve classification across huge volumes of investigative evidence. TNO itself is also testing GPT-NL internally for sensitive projects where commercial AI systems may present privacy or security concerns.

The most striking thing about GPT-NL may be how it was trained. While major AI companies face growing legal battles over copyrighted training data, GPT-NL has reached licensing agreements with Dutch news publishers covering newspapers, broadcasters, and online media platforms across the country. According to the project, it is the first AI initiative anywhere in the world to secure paid collective agreements with all major publishers in a single market.

Sign up for breaking news, reviews, opinion, top tech deals, and more.

That achievement matters because the relationship between journalism and AI companies has become increasingly hostile. Publishers argue that their work has been scraped without permission and repurposed into systems that can compete directly with the original reporting.

GPT-NL’s licensing terms are publicly documented, publishers are compensated, and technical safeguards are intended to prevent users from extracting licensed content directly through prompting.

Still, the project faces the same reality of the cost of scaling that confronts nearly every AI initiative. GPT-NL’s 25 developers and budget are tiny by AI standards. That creates a tension running beneath the optimism surrounding the project. GPT-NL appears usable for institutional use, but continuing to improve the model while keeping pace with global AI development will require sustained funding and political support.

Still, there are only a few meaningful AI challengers outside of the biggest American companies. The Netherlands is effectively testing whether there is another route forward, one centered on public institutions, negotiated copyright agreements, and local control over data infrastructure.