Impact of replacing human participants with language model agents on Condorcet jury theorem and cognitive diversity

Ascertain how substituting human citizens with language model agents acting as digital representatives affects the applicability and outcomes of the Condorcet jury theorem in democratic decision-making, and determine whether such agents preserve cognitive diversity comparable to the human populations they represent.

Background

The paper evaluates proposals to use LLM agents as digital representatives in democratic procedures and raises concerns about the epistemic foundations of democracy. The Condorcet jury theorem and cognitive diversity among participants underpin democracy’s epistemic advantages.

The authors explicitly state uncertainty about whether replacing people with LLM agents would maintain the conditions for the Condorcet jury theorem and preserve cognitive diversity. Resolving this question is crucial for assessing the epistemic legitimacy of delegation to AI agents in collective decision-making.

References

We probably also shouldn't take for granted that this delegation to proxy agents would preserve the epistemic benefits of democracy - we are not sure how the Condorcet jury theorem is affected by having LLM Agents replace the people they represent or if the resulting agents would really have the same overarching cognitive diversity as the people they represent [56, 57], but it does seem that the adversarial political process helps surface information relevant to collective decision-making that would otherwise remain submerged; and by (at least some people) actively participating in public life, they are inspired to come up with novel policy proposals that don't just triangulate among the things that people already believe, but offer genuinely innovative solutions to pressing practical problems.

— Can LLMs advance democratic values?  (2410.08418 - Lazar et al., 2024) in Section IV (Scoring LLMs for Democracy)