Participatieve methoden voor alignment van taalmodellen

Together with TNO, Waag Futurelab is investigating which participatory methods for value alignment exist in the development and use of large language models. In collaboration with civil society organisations, we are testing a method for identifying and embedding values in GPT-NL.

How do we know that generative AI acts in accordance with the values and interests of its users? Whose interest and values should be taken into account when designing large language models? And how can the public contribute and control the development of LLMs? These questions are part of the so-called value-alignment debate of AI. The aim of alignment is for an AI system to behave in line with a shared set of values. But there is no consensus yet on what good alignment looks like and what participation methods are well-suited to give the public a voice in shaping generative AI.

Together with TNO, Waag Futurelab is therefore investigating which participatory alignment methods exist in the development and use of large language models. Language models can be misaligned in many ways, be it by providing incorrect anwers or by creating biased, discriminating, or harmful outputs. Participation is therefore crucial to understand and embed the diverse perspectives and experiences of groups otherwise facing harms.

In collaboration with TNO and different civil society organisations, Waag is exploring existing participatory methods and test one of them for identifying and embedding values in GPT-NL. Developed by TNO, SURF and NFI, GPT-NL wants to offer a responsible alternative to existing language models. The language model is based on lawfully obtained, high-quality, Dutch data. The creators are transparent about which training data has been used and ensure that a portion of the proceeds flows back to the copyright holders. In this way, GPT-NL wants to provide an alternative way for developing European language models and ensures that copyright holders are given a fair place in the development of this technology.

The project will provide concrete recommendations and an overview of participatory methods for the further development and use of GPT-NL.