"Generative Agent Simulations of 1,000 People"

Two-hour interviews let AI clone your decision-making style with 85% accuracy

AI agents built from real interviews predict human behavior as accurately as humans predict themselves

This paper introduces a novel way to create accurate AI simulations of real people using two-hour interviews and LLMs. The system successfully replicates individual attitudes, behaviors, and personality traits across various social science measures, achieving 85% accuracy compared to how consistently people replicate their own responses.

-----

https://arxiv.org/abs/2411.10109

🤔 **Original Problem**:

Traditional human behavior simulations rely on oversimplified demographic stereotypes or manually specified behaviors, limiting their ability to capture real human complexity across different contexts.

-----

🔧 **Solution in this Paper**:

→ The researchers created generative agents for 1,052 real individuals using two-hour qualitative interviews.

→ An AI interviewer conducted semi-structured interviews exploring life stories, views, and experiences.

→ The system injects complete interview transcripts into LLM prompts to simulate individual responses.

→ For multi-step decisions, agents maintain memory through text descriptions of previous interactions.

-----

💡 **Key Insights**:

→ Interview-based agents outperform demographic and persona-based approaches by 14-15 points

→ Even with 80% of interview content removed, agents still maintain strong predictive power

→ The method reduces accuracy biases across racial and ideological groups

→ Two-hour interviews capture more nuanced behavioral factors than traditional surveys

-----

📊 **Results**:

→ 85% normalized accuracy on General Social Survey responses

→ 80% normalized correlation on Big Five personality traits

→ 66% normalized correlation on economic game behaviors

→ Successfully replicated 4 out of 5 social science experiments