The KB, National Library of the Netherlands has entered into a collaboration agreement with GPT-NL; the sovereign Dutch AI language model that is currently being developed by TNO, SURF and the NFI. The KB is one of the ‘launching partners’ of the project and has provided copyright free data for further development of the model. In June, both parties started a study that, among other things, is meant to provide insight into possible use cases of GPT-NL for the KB.
The KB has been on GPT-NL’s ‘content board’ since 2024. This means that the National Library has provided public domain content with GPT-NL for development of the language model. Specifically, it has supplied copyright free, Dutch written language heritage from the KB’s collections. In return, the KB has a say in the development and future of the model.
“We find it important to have a voice as the National Library for two reasons,” explains Niek Schroten, CIO of the KB. “We welcome the creation of a responsible Dutch alternative to the existing commercial parties. We need an AI model for the Dutch language and context, which will supply reliable information to users based on high quality data. As the KB, we happily contribute to GPT-NL’s intention to an eminent and pluralistic data-ecosystem.
According to Schroten, the development of responsible AI is even essential for increasing democratic resilience. “We need free access to reliable information to guarantee freedom of opinion. With AI developing at breakneck speed, we must actively maintain the reliability of training data.”
The second reason for the KB’s collaboration with GPT-NL is because both parties attach great importance to the use of legitimately obtained data. Schroten: “As the National Library, which manages written heritage by Dutch creators, we find it important that copyright owners get a fair compensation for their work. Sources must remain traceable and copyright owners must be treated fairly in the development of new technology. Well-known commercial AI tools have made unauthorised use of copyrighted data. We are against that.”
The study that started this month is a feasibility study that should supply insight into the application of GPT-NL within the KB’s collection search systems. It will investigate where an RAG-assistant (Retrieval Augmented Generation) can better serve users with their search questions within a specific part of the KB collection. This study will use a copyright free dataset (material that is more than 140 years old) from the KB platform Delpher.nl. The study will utilise the super computer Snellius and will run until mid-November of this year.
Schroten: “We purposely chose to use copyright free data from our collection of newspapers, books and magazines for the feasibility study. We intend for the study to be beneficial for both sides. The results of the study will teach GPT-NL more about the type of answers that her language model gives, and the KB will lear more about possibilities in future services.”
The first version of GPT-NL, GPT-NL 1.0, was soft-launched in February 2026. This is a beta version that is currently being tested by GPT-NL’s launching partners. Besides the feasibility study that is now being run with the KB, GPT-NL is running similar studies with three other launching partners: the Ministry of the Interior, TNO and the NFI. The launch of the public version of GPT-NL is expected at the end of the year.