AI can imitate the average person, but real people disagree and shift their answers. Can a simulation fake that mess?

This explores how the natural spread in human answers, both between different people and from the same person, affects whether an LLM simulation can be trusted to stand in for real people. The corpus suggests variation is the thing simulations most often get wrong, and that the usual accuracy checks tend to hide the problem.

The clearest failure is that models are too uniform. When researchers compared 106 LLMs against human respondents on 625 value scenarios, the models clustered in a narrow, idealized region while people scattered widely (Do large language models actually reflect human value diversity?). The same pattern shows up when a simulation is built from persona prompts. Running one persona prompt repeatedly produced as much variation between runs as switching between different personas did (Why do LLM persona prompts produce inconsistent outputs across runs?). So the variation you see in the output is mostly model uncertainty, not knowledge of how different kinds of people differ. That makes persona prompting a poor way to reproduce something like annotator disagreement.

Training is part of the cause. Pretrained base models conditioned on short dialog samples predicted human responses more accurately and more diversely than instruction-tuned assistants prompted with personas (Do pretrained models simulate humans better than instruction-tuned assistants?). Tuning a model to be a helpful assistant appears to narrow its range, which works against imitating people. In the other direction, models finetuned on psychology experiment data captured individual differences in their embeddings and beat traditional cognitive models at predicting decisions (Can language models learn to model human decision making?). Variation can be learned if the model is trained on data that contains it.

Some methods build variation in on purpose. RecLLM conditions a user simulator on session-level profiles and turn-level intents, which makes the synthetic conversations harder to tell from real ones (Can controlled latent variables make LLM user simulators realistic?). The control has limits, though. Simulators can lose track of their own goals over a multi-turn conversation, and that drift corrupts the training signal for whatever learns from them (Why do LLM user simulators fail to track their own goals?). Varied users are only useful if each one stays consistent.

The corpus also says two ways of checking accuracy can mislead. Setting temperature to zero makes outputs repeatable, but a repeated output is still one draw from a distribution, so consistency is not reliability (Does setting temperature to zero actually make LLM outputs reliable?). Matching aggregate behavior is not enough either, because behavior underdetermines the reasoning behind it (Can matching human actions prove an LLM simulation explains behavior?). Even the good replication result, 76 percent of published marketing-experiment main effects, was strongest for effects that were already statistically strong and unreliable for marginal ones (Can AI personas reliably replicate human experiment results?). Marginal effects are where human variation matters most, since they are small differences buried in noise. The practical test is whether a simulation reproduces the spread of human responses, not only their average.