Google Researchers Can Create an AI That Thinks a Lot Like You After Just a Two-Hour Interview

Abstraction: Stanford/Google study simulates 1,000 people as LLM agents from interview transcripts

Key points:

• Stanford + Google DeepMind study "Generative Agent Simulations of 1,000 People" (arxiv 2411.10109): 1,052 participants paid $60 for a 2-hour AI interview; transcripts (~6,491 words each) fed into LLMs to create persona agents
• Agents matched real participants on General Social Survey with ~85% accuracy and Big Five Personality Inventory at ~80%
• Performance dropped to ~66% normalized correlation on economic games (Prisoner's Dilemma, Dictator's Game) where real cash incentives applied
• Pitched as tool for policymakers and businesses to simulate public response to policies, products, or shocks without repeated polling
• Raises misuse concerns: same tech could enable voice-cloning scams or corporations/politicians claiming mandate from synthetic approximations of public will

Connections: Google · Stanford University · AI Agents · Large Language Models

Source: https://gizmodo.com/google-researchers-can-create-an-ai-that-thinks-a-lot-like-you-after-just-a-two-hour-interview-2000547704