How AI simulation could model human behavior

In this podcast conversation, the hosts interview June, a Simile cofounder whose research helped establish generative agents as a way to model social behavior. He argues that foundation models can be trained on real decisions to reproduce human choices—including their errors and inefficiencies—rather than idealized answers. The discussion moves from causal evidence and validation limits to multi-agent societies, cost, and scale. It ends with the harder promise: using simulations to represent people who cannot be consulted every time a consequential decision is made.

Web-trained foundation models contain traces of human behavior that can support richer social simulations. Reproducing behavior at scale may require changing a model's parameters, rather than retrieving memories or writing better prompts.

• Broad training lets foundation models produce human-like behavior that narrow systems cannot.
• Web-scale text provides an incomplete but unusually broad record of how people behave.
• A ten-year thought experiment led the researchers to social simulation as a consequential use of the technology.
• Memory retrieval differs from learning deeper behavioral patterns that may need to be encoded in model parameters.

Predicting an outcome does not show how to change it. Simulations become useful when randomized evidence helps estimate which intervention could move a complex system toward a chosen goal, and by how much.

• Forecasting a bad outcome is different from identifying an action that could prevent it.
• The world produces only one realized outcome, so ordinary observations rarely reveal what a different choice would have caused.
• Randomized experiments with real stakes provide evidence about how interventions change decisions.
• For politically sensitive or complex decisions, the size and precision of an effect matter alongside its direction.

The target is the choices people actually make, including biases, inconsistencies, and meaningful inefficiencies—not an idealized rational agent. Validation compares simulated behavior with real decisions while accounting for weak studies, selective publication, and uneven performance across populations.

• A behavior model should reproduce people's mistakes instead of returning a super-rational answer.
• Performance can vary substantially for niche populations, limiting claims that a system represents everyone equally well.
• The evidence base inherits social science's replication problems and its tendency to publish statistically significant results.
• Useful simulations must preserve ordinary preferences, such as choosing a walk that helps someone think even when it is not the fastest route.

The ambition is to place behavior models inside multi-agent environments where individual choices produce collective patterns over time. More data and compute may improve the models, while their cost should be compared with real-world studies and decisions.

• Early results suggest that additional human data and compute improve predictive performance.
• The longer-term plan places learned agents in richer environments where they can interact instead of answering isolated questions.
• Classic agent-based models show how small individual preferences can create large collective patterns, but they describe people too simply.
• The relevant comparison is the cost of simulating thousands of people against the expense and consequence of studying or acting in the real world.

Simulation could represent people who cannot be repeatedly surveyed whenever a decision affects them. The analogy to painting captures the design challenge: a useful representation must preserve mundane details that reveal something deeper about individuals and society.

• Market research addresses only part of the broader problem of understanding human decisions.
• Organizations cannot practically survey every affected stakeholder about every choice they make.
• Simulated behavior could represent people's voices where consequential decisions are made.
• Like painting, simulation can use ordinary details to reveal what is essential about a person or society.