An AI persona can nail a single survey answer, so why does it still lose track of who it's supposed to be?

This explores the difference between a simulated persona matching what a real person would say on a single question or experiment, and that persona staying the same character across a long conversation. The corpus treats these as two separate tests with different failure modes. Snapshot fidelity asks whether the answer lands near the human answer. Sustained consistency asks whether the same someone keeps answering. A persona can pass the first and fail the second.

Snapshot fidelity looks good on paper, but the numbers are strongest where the stakes are lowest. AI personas reproduced 84 of 111 published marketing-experiment effects, and success tracked how strong the original evidence was (Can AI personas reliably replicate human experiment results?). Personas built from real behavioral data called A/B test directions correctly 75 to 90 percent of the time, but they were least trustworthy when the true effect was near zero (Can behavior-based personas predict A/B test outcomes?). The broader picture is that headline accuracy of 76 to 85 percent hides run-to-run instability, resistance to personality conditioning, and cognitive biases that match the persona's identity (How accurately can language models simulate human personalities?). Run-to-run instability is already a hint of the consistency problem. If the same persona gives different answers on different runs, a good average is masking a wobbly character.

Sustained consistency fails in different ways, and researchers had to split it into separate failures to measure it. One line of work distinguishes local drift within a turn, global drift across a conversation, and outright factual self-contradiction. It trains user simulators against all three with multi-turn RL and cuts drift by over 55 percent (Can training user simulators reduce persona drift in dialogue?). Short character descriptions don't hold up over a dialogue. A layered architecture that confines the model to reactive replies inside a structured script produces more human-like consistency, but at the cost of emotional range (Can layered persona architecture sustain coherent character behavior?). The metric itself is also easy to fool. High consistency scores often come from parroting the character sheet while ignoring what was just asked, so persona and discourse coherence have to be optimized together (Do persona consistency metrics actually measure dialogue quality?). Fidelity is also multidimensional, and judging it in parts beats a single overall verdict (Can breaking persona fidelity into parts improve how we judge it?).

A deeper reason for the gap is where the persona lives. Persona prompts steer the output channel. They make a model follow trait instructions, but the underlying between-group bias stays put and is redistributed rather than removed (Can persona prompts actually reduce bias in language models?). Contrast that with the model's own trained persona. A separate strand argues that post-training installs dispositions that survive adversarial pressure, while prompt-induced role-play collapses under jailbreaks (Are RLHF personas performed characters or realized dispositions?, Are LLM personas realized or merely simulated through training?). Read together, they suggest a simulated persona is a thin costume over a sticky trained character. A costume can match a single answer without being anchored to anything that keeps it steady over time. That is an inference from these notes and not something they test directly.

The corpus does not answer the sharpest version of your question. The snapshot studies measure agreement with human answers, and the drift studies measure dialogue self-consistency. No note here checks whether a persona that replicates survey or experiment results also keeps its values through a long conversation. Until someone runs that test, high single-answer fidelity should not be treated as evidence of a stable character.