Memetic Drift in Multi-Agent LLMs: Scaling Laws for Consensus Under Pluralistic Uncertainty

Pluralistic alignment often requires AI systems to aggregate or deliberate over multiple defensible perspectives, but consensus alone does not reveal whether an outcome reflects collective reasoning, systematic bias, or chance. We study this ambiguity in near-tie discrete-choice settings, where several options are defensible and no external reward, ground truth, or payoff initially selects among them. We introduce Quantized Simplex Gossip (QSG), a minimal null model in which agents maintain simplex-valued belief states, communicate quantized samples, and adapt locally to one another's outputs; QSG traces consensus formation to mutual in-context learning, a regime we call memetic drift. The model yields early-drift identities and mean-field scaling laws in population size, communication bandwidth, adaptation strength, and internal uncertainty, and predicts a drift--selection crossover under weak asymmetries. We evaluate these predictions in a contextualized municipal budget-plan prompt with neutral plan codes, neutral naming games with GPT-4o and Claude Haiku 4.5, and QSG simulations, providing a diagnostic baseline for distinguishing evidence aggregation from amplified sampling noise.