Canvis a "Decidim to democratize AI models and services"

Proposed agenda item

We should actively promote and experiment with Decidim as a interface to design, build, and govern AI models and services.

Background

We need to democratize AI urgently. I think AI is the most powerful technology ever created and that it captures and automatizes the general intellect. Unfortunately, this capture is mostly privatized and instrumentalized for marketing and military purposes. But there are some exceptions. A few public institutions are developing public open-weight generative AI models. But these have no democratic participation, other than the indirect contribution we all provide through the training data.

Among the short-time opportunities in this direction: The Spanish project ALIA has developed a series of AI models called Salamandra that work pretty well. But they are raw foundation models: they have not been ethically and politically trained. This is an open opportunity for participatory democratization. We want to propose that Decidim be used as an interface for this democratic fine-tuning. More specifically, the easiest way to contribute to this task is to use Decidim to write a “constitution” for AI. This is a specific technique meant to create a set of principles, like a legal constitution, to govern the behavior of the generative AI model. Once the constitution is written, another artificial intelligence teaches the target system to behave according to the constitution. Other public and open/free models like the Basque Latxa, are also in need and interest to interface democracy.

Intended outcome(s)

I expect the Committee and the community to take this as a possibility and a challenge to be pursued (not a primary goal of the association, nor major resource investment)