The KB makes royalty-free collections available for Dutch AI model

Generative Artificial Intelligence (AI) models such as ChatGPT need to be trained with existing data such as text or images. The KB has a large quantity of digital texts in its collections that will now start contributing to an ethically responsible AI language model. As part of this, the KB has signed an agreement with GPT-NL, the partnership developing this Dutch language model.

For society to work with artificial intelligence in an ethically responsible way, it is essential to invest in public alternatives to AI models such as ChatGPT and Gemini. This will reduce our dependence on commercial parties and give us as a society more control over how technology is developed and deployed.

Commercial companies train AI models with large amounts of text that are not legally obtained. We also do not know what Dutch data was used, and it is quite possible that mainly English or translated text was used as training data. These AI models therefore lose the nuances of Dutch language and culture that are necessary for some applications.

The KB embraces digital developments and at the same time considers it crucial for AI companies and other organisations to operate ethically. For example, it is essential that copyrights are respected, sources are properly cited and personal data is protected. 

The GPT-NL project is an important response to these commercial AI models. GPT-NL is using only lawfully obtained data to develop an ethically responsible AI language model for Dutch language and culture. The KB stands fully behind the development of GPT-NL, precisely because the project explicitly addresses responsible data use, fair agreements with copyright holders and the careful handling of cultural heritage. 

We support this initiative and contribute to it by offering the use of all royalty-free digital texts from our collections. GPT-NL is currently being trained with this data and the data made available by other parties, including ICTRecht, Het Utrechts Archief and De Nederlandsche Bank. An initial version of GPT-NL will be available by the end of 2025.