AI Democratization

What do people mean when they talk about “AI democratization”?

AI companies around the world – e.g. Stability AI, Meta, Microsoft, Hugging Face, H20.ai – are talking about their goals for “democratizing AI”, but it’s not always clear what they mean. The term “AI democratization” seems to be employed in a variety of ways causing commentators to speak past one another.

Colleagues and I have outlined at least four concepts of “AI democratization in use:

  1. The democratization of AI use is about enabling a wide range of people to use and benefit from AI applications.
  2. The democratization of AI development is about helping a wider range of people contribute to AI design and development processes.
  3. The democratization of AI benefits or profits is about facilitating the broad and equitable distribution of value accrued to organizations that build and control advanced AI capabilities.
  4. The democratization of AI governance is about distributing influence over decisions about AI to a wider community of stakeholders and impacted populations. AI governance decisions involve balancing AI related risks and benefits to determine how and by whom AI is used, distributed, developed, and regulated.

AI Democratization is not the same as model dissemination

Very often AI Democratization is discussed in the context of model sharing decisions – that is, when developers are deciding, for instance, whether to open-source an AI model (generally, to make its weights and training code available for anyone to download, use, and modify) or to make the model accessible to users or for research purposes via API.

Model sharing, and particularly open-sourcing, promotes the democratization of AI development and the democratization of AI profits (largely via the democratization of development). However, open model sharing should not be conflated with AI democratization.

First, AI democratization is a multifaceted project which, in addition to model sharing, can also benefit from proactive efforts to help a wider array of people benefit from AI applications and influence AI futures. These might include the provisions of educational and up-skilling opportunities, intuitive user interfaces, project support and coordination, profit redistribution schemes, fundamental technical infrastructure, and opportunities to influence impactful decision-making about AI (e.g. through multistakeholder bodies, participatory processes, or representative deliberations).

Second, the goals of different forms of AI democratization can sometimes conflicts. For example, if the public prefers for access to certain kinds of AI systems to be restricted, then the “democratization of AI governance” may require access restrictions to be put in place—but enacting these restrictions may hinder the “democratization of AI development” for which some degree of AI model accessibility is key.

For more on the risks and benefits of open-sourcing increasingly capable AI systems, see the open-source model sharing project page.

Further resources on AI Democratization

Seger, E., et al. (2023). Democratising AI: Multiple Meanings, Methods, and Goals. Proceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society

Seger, E. (2023). What do we mean when we about when we talk about “AI Democratisation? GovAI Research Blogpost

The Collective Intelligence Project (CIP) is a research incubator for democratic AI governance models. In particular, see their recent work on alignment assemblies.