Public Constitutional AI

Abstract

We are increasingly subjected to the power of AI authorities. Machine learning models now underpin algorithmic markets, determine whose speech is amplified or restricted, shape government decisions ranging from resource allocation to predictive policing, and influence our access to information on critical issues such as voting and public health. As AI decisions become inescapable, entering domains such as healthcare, education and law, we must confront a vital question: how can we ensure that AI systems, which increasingly regulate our lives and make decisions that shape our societies, have the authority and legitimacy necessary for effective governance?

To secure AI legitimacy, this essay argues, we need to develop methods that engage the public in the project of designing and constraining AI systems, thereby ensuring that these technologies reflect the shared values and political will of the community it serves. Constitutional AI, proposed and developed by Anthropic AI, represents a step towards this goal, offering a model for how AI might be brought under democratic control and made answerable to the common good.

Just as constitutions limit and guide the exercise of governmental power, Constitutional AI seeks to hardcode explicit principles and values into AI models, rendering their decision-making more transparent and accountable. What sets Constitutional AI apart is its commitment to grounding AI training in a clear, human-understandable "constitution." By training AI to adhere to principles that are legible to both humans and machines, this approachaims to foster trust and stability in the development of these increasingly powerful technologies.

However, I argue that Constitutional AI, in its current form (developed by a private corporation seeking to create universally applicable constitutional principles), is unlikely to fully resolve the crisis of AI legitimacy due to two key deficits. First, the opacity deficit: the inherent complexity of AI systems undermines our ability to reason about their decision-making. Second, the political community deficit: AI systems are grounded in abstract models rather than human judgment, lacking the social context that legitimizes authority.

To remedy these deficits, I propose Public Constitutional AI, a framework that involves the public in drafting an AI constitution that must be used in the training of all frontier AI models operating within a given jurisdiction. By transforming the AI constitution from a technical solution devised by engineers into a product of significant citizen involvement, Public Constitutional AI mitigates the opacity deficit. It does so by rendering the principles and values governing AI systems more transparent and accessible to the forms of public discourse and contestation that are essential to democratic legitimacy. Moreover, by grounding the development of AI principles in the social context and shared experiences of a particular political community, Public Constitutional AI helps to bridge the gap between the abstract logic of algorithms and the situated, contextual judgments that legitimize authority in a democracy, thereby mitigating the political community deficit.

Keywords: Public Constitutional AI, Ai Governance, Democratic Legitimacy, Algorithmic Decisiom-making, AI ethics, Digital Constitutionalism, Automated Authorities, AI Legitimacy, Technological Due Process, AI Courts