I present evidence that AI reasoning models develop internal "parliaments", recurring argumentative voices with distinct behavioural profiles, when making decisions under uncertainty in game-theoretic environments. Using an automated Deliberation Detector applied to full chain-of-thought traces from models across six AI laboratories (Anthropic, Alibaba, Google, DeepSeek, xAI, and OpenAI), I identify six voice archetypes (Analytical, Conservative, Aggressive, Contrarian, Intuitive, Neutral) and quantify their frequency, win rates, and correlation with decision quality. Key findings: (1) reasoning models debate internally in 7–53% of decisions depending on laboratory of origin, (2) only 1–21% of these debates reach explicit resolution: the vast majority are performative, (3) a single dominant voice wins 90%+ of debates regardless of model size, (4) when the Analytical voice overrides the default, decision quality improves dramatically, (5) different AI laboratories produce fundamentally different parliament structures, and (6) one major provider charges for reasoning tokens but does not expose the reasoning text.
Venelin Videnov (Wed,) studied this question.