Large dataset news organizations for Dutch AI language model GPT-NL

The collaboration between NDP Nieuwsmedia and TNO supports the shared goal of GPT-NL and the Dutch government: to create a language model that respects copyright and sets an benchmark for how to handle copyrighted content in AI systems. Strict agreements have been made to prevent the articles from being technically extracted from the AI-model. When GPT-NL is released, publishers will receive appropriate remuneration.

GPT-NL is a non-profit initiative by TNO, NFI, and SURF, and offers a responsible alternative to existing language models. It is built for the Netherlands using high-quality Dutch data. Unlike some international models that use large parts of content scraped from the internet without permission, GPT-NL collects copyrighted data carefully and ethically, and remunerates those who contribute their content. It also complies with European laws like the AI Act. The model is being developed for specific tasks such as summarizing, simplifying, and extracting information from text.

News article data

Thanks to the collaboration with NDP Nieuwsmedia—the trade association of private news publishers such as DPG Media, Mediahuis, Erdee Mediagroep, and De Groene Amsterdammer, and the participation of ANP, the model gains access to over 20 billion tokens. The articles cover many topics, from politics and economics to healthcare and science, and form a rich source for training GPT-NL. The datasets contain billions of tokens, small pieces of text that help AI understand and process language. A token can be a word, part of a word, or even a punctuation mark.

Timeline

TNO, NFI, and SURF began developing GPT-NL in 2023. Training started in June 2025. In the fourth quarter of 2025, the model will be improved and prepared for first use. Earlier Dutch contributors of data to GPT-NL include DNB (De Nederlandsche Bank), ICTRecht, and Het Utrechts Archief.