Feed an AI someone's full profile and it still can't predict what that person will do, even when crowds look right.

This explores why giving an LLM a person's profile doesn't make it better at predicting what that particular person will do, even though persona simulations often get group-level results right. The most direct evidence comes from a test across 208,021 participants. Conditioning models on each participant's profile produced no meaningful gain in predicting those specific individuals Does conditioning LLMs on personal profiles improve prediction?. The standard technique for turning a generic model into a particular person doesn't measurably do that.

The failure is easy to miss because persona simulation looks good in aggregate. AI personas reproduced 76 percent of published marketing-experiment main effects, and their success tracked how strong the original evidence was. The marginal effects were unreliable Can AI personas reliably replicate human experiment results?. Personas built from real behavioral data called the direction of A/B tests correctly 75 to 90 percent of the time. They did best on large effects and worst on near-zero ones Can behavior-based personas predict A/B test outcomes?. Both tests ask whether the crowd tips the right way, and neither asks whether Person A differs from Person B. A model can pass them without capturing any individual variation, and individual variation lives in the small differences.

The corpus offers three clues about why the persona doesn't individuate. First, personas work at the surface. Persona prompts made models follow trait instructions, but the bias gaps between groups stayed unchanged, so the prompt changes what gets said rather than what's underneath Can persona prompts actually reduce bias in language models?. Second, when personality-conditioned agents lived through simulated life events, they did change, but generically. The shifts were weakly tied to the specific event, smaller than human effect sizes, and squeezed together across personas, so each persona moved toward a common pattern instead of along its own path Do personality-conditioned agents change like humans do?. Third, post-training installs stable dispositions that hold up under adversarial pressure Are LLM personas realized or merely simulated through training? Are RLHF personas performed characters or realized dispositions?. The corpus doesn't test this directly, but a plausible reading is that a persona prompt is a thin layer over a deeply trained default. That default would pull every persona toward the same center.

A related failure may also be at work. LLMs look socially competent when one model controls everyone in the scene, but they fail systematically once agents hold private information Why do LLMs fail when simulating agents with private information?. Much of what makes one person behave differently from another is private, such as their history, circumstances and current situation. A short profile can't carry most of that, and the model has no grounding to fill the gap.

The most promising hint is PersonaAgent, which treats the persona as something that evolves. It rewrites the persona at test time by simulating recent interactions against feedback, and the learned personas cluster by user in latent space Can personas evolve in real time to match what users actually want?. That suggests behavioral traces carry what a static description doesn't, though the summary doesn't show it beating baselines at individual prediction. Don't confuse this failure with persona drift. Multi-turn RL cuts a simulator's drift by 55 percent, but that only keeps it in character across a conversation Can training user simulators reduce persona drift in dialogue?. A persona can be perfectly consistent and still be wrong about the person. The corpus documents the failure well but doesn't isolate one cause, so the mechanisms above are converging evidence, not a settled answer.