If an AI is built from your interview, can it predict what you'd say next — about as well as you can?
How accurate are interview-trained agents at predicting what people would actually say?
This explores how closely AI agents built from in-depth interviews with a person can forecast what that person would say when asked new questions, and how far that number can be trusted.
This explores how closely AI agents built from in-depth interviews can forecast what the interviewed person would say to new questions. The best number in the corpus is about 85%. In a study of 1,052 people, agents built from voice interviews reproduced participants' survey-style responses Can AI agents learn people better from interviews than surveys? about 85% as accurately as the participants reproduce their own answers when asked again later. The benchmark is human self-consistency, not perfection. People contradict themselves over time, so the agent is close to the ceiling of what any predictor could do.
What drives the accuracy is somewhat surprising. Factual content mattered more than linguistic style. Even a bullet-point summary of the interview kept 83% fidelity Can AI agents learn people better from interviews than surveys?. The agent seems to work because it knows what you have said and done, not because it imitates how you talk. A companion finding from the model side points the same way: pretrained base models given short samples of real dialogue predicted human responses more accurately and more diversely than instruction-tuned assistants prompted with a persona description Do pretrained models simulate humans better than instruction-tuned assistants?. Real material about a person beats a description of one, and the assistant training that makes chatbots helpful gets in the way of simulating humans.
The corpus also suggests limits on how far the 85% stretches. RLHF-trained models tend to predict conciliatory, polite persuasion whatever the actual dialogue looks like, projecting their own accommodating training onto other people Do LLMs predict persuasion based on actual dialogue or training bias?. When persona-conditioned agents live through life events, they shift in generic ways that are smaller than human changes and weakly tied to the person or the event Do personality-conditioned agents change like humans do?. An interview agent is a snapshot of someone at one moment. It may be much weaker at predicting how they would react to a new situation, or how they would change.
There is also a difference between predicting and being. Models can beat every individual human at predicting social norms, yet they cannot take part in the community processes that create those norms Can AI predict social norms better than humans?. Shanahan's view of dialogue agents fits here. An interview agent plays a character assembled from what it was told, with no underlying person Does a language model have an authentic voice underneath?. That may not matter for forecasting survey answers, but it matters if you treat the agent as a stand-in for the person. In a related finding, small models trained to abstain when uncertain matched models ten times larger at conversation forecasting Can models learn to abstain when uncertain about predictions?. So a useful interview agent should also know when its prediction is a guess.
The corpus does not cover how these agents do on real behavior outside survey-style answers, or on questions the interview never touched. The 85% describes agreement on answers, not accuracy about people's lives.