Study: LLMs show one-third the perspective diversity of human groups

A new paper by Maurice Flechtner, 'The Deliberative Deficit: An Empirical Critique of LLMs in Democratic Discourse' (arXiv:2608.10186), challenges the assumption that strong performance on verifiable benchmarks translates to sound collective reasoning on value-laden problems. Deploying LLMs in citizen assemblies, policy forums, or ethics committees requires integrating pluralistic perspectives, yet most evaluations only test math, coding, or coordination games. Flechtner applies the Deliberative Reason Index (DRI), a validated measure from political science, to assess reliable group-level reasoning across 1,980 five-agent LLM runs on 12 citizen-assembly topics and 11 frontier model configurations.

Results show LLM groups achieve procedural quality comparable to human deliberation—respectfulness, justification, and engagement—but intersubjective consistency gains are small, topic-dependent, and concentrated on tractable questions rather than ethically contested ones. Critically, LLM groups exhibit one-third the perspective diversity of human assemblies, and unlike humans—who converge as diverse views synthesize—LLM deliberation increases opinion dispersion. Engineering diversity through persona prompting doesn't restore human dynamics; it simply shifts which deliberative component gets updated. Flechtner's conclusion is constraining rather than prohibitive: LLMs can serve as tools for supporting human reasoning but current evidence does not license treating them as autonomous deliberative agents.