If you give an AI a richer backstory, does it finally choose the way the real person would?

This explores whether richer persona descriptions, going beyond demographic labels, can make AI agents choose the way their human counterparts actually do. The corpus doesn't contain a study that directly tests this. But the pieces it does have point to an uncomfortable answer: the gap may sit beneath the prompt, where more prompt detail can't reach it.

Start with the clearest case of the gap. When ChatGPT agents were given the demographic profiles of real study participants, they chose truthful answers much more often than those people did. The humans preferred confident falsehoods to answers that admitted uncertainty, and the agents didn't follow Can demographic prompts make ChatGPT replicate human preferences?. Read naively, that looks like missing information: tell the model more about who the person is and it will act more like them. A separate finding complicates this. Persona prompts change what a model says on the surface, but the underlying bias stays where it was. Sentiment gaps between groups persist even when the model obeys trait instructions Can persona prompts actually reduce bias in language models?. If persona conditioning only acts at the output level, a richer persona is still a costume over the same underlying values. A model trained to favour honesty may keep favouring honesty however vividly you describe someone who doesn't.

A useful way to see why comes from treating dialogue agents as role-playing characters. The prompt sets up a character, and the model writes text that fits it Should we treat dialogue agents as role-playing characters?. Role-play is good at producing what a character would plausibly say. It is less good at reproducing what a real person would actually choose when no one is checking, especially when that choice runs against what the model was trained to see as a good answer. Preferring a confident lie is exactly that kind of choice.

Richer personas do measurably help with a different problem: consistency. A three-layer 'Deep Persona' design that keeps the model responding within a structured script produced dialogue closer to human conversation patterns than shallow character descriptions did, though emotional expression stayed weak Can layered persona architecture sustain coherent character behavior?. Training user simulators with multi-turn reinforcement learning cut persona drift by over 55% Can training user simulators reduce persona drift in dialogue?. Notice which fix did the heavier lifting: changing the model through training, not just enriching the prompt. That hints that closing a preference gap, not just a consistency gap, may also need training-level work.

The most practical calibration comes from the marketing replication work. AI personas reproduced 76% of published experimental effects, but success tracked how strong the original effect was. Robust effects replicated, while marginal ones produced both false positives and false negatives Can AI personas reliably replicate human experiment results?. So the agent-human gap isn't spread evenly. It is widest where human behaviour is subtle, contrarian or socially awkward, which is also where a model's trained-in values are most likely to override any persona you give it. What you probably didn't know you wanted to know: a persona that looks more realistic can make a simulation seem trustworthy without making its choices any more human.