Motivation

Why is democratic AI important?

First, human agency: as AI systems increasingly shape our world and become centers of power, we need a way to ensure that meaningful human agency is embedded in decision-making about AI. Democratic systems, by definition, provide a means for enabling intentional human agency over such decisions. Second, pragmatically, well designed democratic systems have useful properties which can help overcome inaction biases and enable the creation of trusted bodies, unlocking critical national and international coordination.

Why did you develop this framework?

There are both normative and pragmatic reasons why it is critical to democratize some of decisions around AI development, deployment and alignment. To do that effectively we need a roadmap or framework to understand where we are and how to move forward. The goal is not to apply democratic processes to every decision but to ensure that the capability exists for critical decisions, and to be able to communicate effectively about when that is and is not the case. Moreover, right now there is extensive discussion about democratizing AI, democratic AI, and democratic input — and much of it is not grounded in meaningful democratic systems.

Democracy

How do you define democracy?

The Democracy Levels Framework builds on the most basic definition of democracy from the Stanford Encyclopedia of Philosophy: “a method of collective decision making characterized by a kind of equality among the participants at an essential stage of the decision-making process.” We are intentionally general and do not assume specific mechanisms, e.g. elections, though we do focus particularly on deliberative democratic processes because of the useful and relatively underrecognized properties that they provide.

Democracy is more than just collective decision making and democratic processes, is that incorporated in this framework?

The word democracy has many connotations in different contexts. For simplicity and having a finite scope for this framework, we currently are only focusing on democratic decision-making processes and institutions that execute them, roughly analogous to legislative decision making and constitutional structures and procedures. (The partial exception to this is in the extrinsic dimensions, which are somewhat impacted by the aspects encompassed by the broader notions of democracy.)

Is this just about deliberative democracy? (What about elections?)

The democracy levels framework is meant to apply regardless of the democratic approach. In the paper, we focus more on deliberative democracy and related approaches in our examples because of the unique properties that make such processes useful for decision-making about AI, particularly for decisions around issues that cross jurisdictional divides, where there is urgent need and the least capability. As we describe in the paper body, “Deliberative democratic processes can work with jurisdictions of arbitrary size, infrastructure, and political structure (including globally), and can effectively incorporate the knowledge of diverse participants and subject matter experts”. Where existing electoral democratic institutions work, and appropriately fulfill the democracy levels, there may be no need for further innovation. Elections also have many useful properties, and especially apply to domains where there is little spillover impact from the AI systems (e.g. within a narrow domain of users or geographies), but are less applicable to e.g. multinational processes, or global regulators.

Is this just about corporations? Alignment? Governments? Regulators? Multinational agreements?

All of the above can be supported by democratic processes and systems, all are in scope.