Democratizing AI Ethics with Collective Constitutional AI

Collective Constitutional AI is an innovative approach that empowers the general public to directly shape the ethical guidelines governing artificial intelligence models. By utilizing Polis-style deliberative polling, this method allows a diverse group of citizens to submit, debate, and democratically vote on the core values an AI should follow. Instead of relying solely on a small group of tech executives and engineers to decide how AI behaves, this process crowdsources a "constitution" of ethical principles that bridges divides and reflects broad public consensus.

Understanding Constitutional AI

To understand the collective approach, it helps to first understand standard Constitutional AI. Typically, AI models are trained to be helpful and harmless using human feedback. In Constitutional AI, the model is given a set of written principles—a constitution—and is trained to critique and revise its own responses based on those rules. Historically, the creators of the AI wrote these rules themselves, raising concerns about hidden biases and a lack of public representation.

The Role of Polis-Style Deliberative Polling

Collective Constitutional AI solves the representation problem by integrating Polis, an open-source, real-time system designed for gathering and analyzing large-scale public opinion. The process works through several distinct steps:

  1. Diverse Recruitment: A representative sample of citizens is recruited, ensuring diversity across age, gender, race, geography, and political ideology.
  2. Rule Submission: Participants are asked to submit written statements detailing the ethical principles they believe the AI should follow. For example, a user might submit, "The AI should choose responses that are objective and neutral."
  3. Voting: As statements are submitted, other participants vote on them by clicking "Agree," "Disagree," or "Pass." To prevent toxic arguments and echo chambers, participants cannot reply directly to one another; they can only vote on existing statements or submit new ones.

Finding Consensus Across Divides

The true power of Polis-style polling lies in how it processes the votes. The platform uses machine learning to map participants into different opinion groups based on how they vote.

Rather than simply picking the rules with the most total votes—which would allow a simple majority to overrule minority groups—the algorithm looks for statements that achieve consensus across different groups. If two deeply divided political groups both agree on a specific ethical principle for the AI, that principle is elevated. This ensures that the resulting rules are not polarizing, but instead represent shared human values.

Training the AI on Public Values

Once the deliberative polling is complete, the statements that achieved the highest cross-group consensus are compiled into a new, publicly sourced constitution.

AI researchers then use this public constitution to train a language model. During training, the AI generates answers to prompts, checks its answers against the public's rules, and rewrites its responses until they align with the democratically chosen principles. Evaluations of models trained this way have shown that they exhibit less bias than models trained solely on rules written by AI developers, while maintaining the same level of performance and safety.

Why Democratic AI Governance Matters

As artificial intelligence becomes deeply integrated into daily life, the values embedded within these systems will impact everyone. Collective Constitutional AI proves that it is possible to include the public in high-level technical governance. By using Polis-style deliberative polling, AI labs can ensure their models are guided by a transparent, democratic, and truly representative set of ethical principles.