AI agents trained on personal interviews can predict individual behavior without task-specific training

Predicting human behavior typically demands outcome-specific training data and substantial work to adapt models to new domains. A new study suggests that large language models might sidestep these constraints by tapping into self-reported information to create generative agents—digital simulations of individuals capable of forecasting responses across a range of scenarios.

Building agents from personal data#

Researchers recruited over 1,000 Americans from a nationally representative sample and constructed digital agents using three approaches: two-hour qualitative interviews conducted through the American Voices Project framework, structured survey responses including personality assessments and standard social science items, or a combination of both. The goal was to test whether these agents could predict how real people would respond to questions they hadn't been explicitly asked during training.

Accuracy rivaling human consistency#

When tested on held-out General Social Survey questions, the agents performed remarkably well. Interview-based agents matched participant responses with 83% of the accuracy that individuals themselves demonstrate when answering the same questions two weeks apart. Survey-trained agents achieved 82%, while agents built on both sources together reached 86%—substantially outperforming agents trained only on basic demographic data, which managed just 74%.

Broad predictive reach and reduced bias#

Beyond standard survey items, the agents also successfully predicted personality traits, choices in economic games, and responses in experimental settings. A notable finding emerged around fairness: agents grounded in rich self-report data—whether interviews or surveys—showed smaller accuracy gaps across racial and ideological groups compared with demographic-only models, suggesting they capture meaningful individual variation rather than relying on demographic proxies.

Diminishing returns from combining sources#

Combining interviews and surveys produced the best results, yet the improvement over using either source alone was modest. This hints that predictive gains from additional data begin to level off once models have absorbed enough contextual information about a person within a given domain.