A team of researchers at Stanford University has successfully replicated the unique thoughts and personalities of over 1,000 individuals using artificial intelligence, sparking both excitement and concern about the ethical implications of such technology. The findings, detailed in a report published in New Scientist, showcase the potential of generative AI to simulate human behavior with remarkable fidelity, while raising questions about the risks of imitating individuals so precisely.
The experiment, led by Joon Sung Park, sought to create AI agents that could accurately model the complexity of individual human behavior. To achieve this, the team recruited 1,052 participants from across the United States, representing a diverse cross-section of the population. Each participant engaged in a two-hour, voice-enabled interview conducted by GPT-4o, the latest iteration of OpenAI’s ChatGPT technology. The AI followed a general script but adapted its questions in real time based on the participants’ responses, probing topics such as personal history and societal views.
The resulting interviews were transcribed and used to train individual AI agents for each participant. These AI agents were then tested on a variety of metrics, including the General Social Survey (GSS), personality assessments, behavioral economic games, and social science experiments. The AI agents’ responses were compared to those of the original participants, who retook the same tests two weeks later.
While human responses naturally varied between the two sessions, with a match rate of 81% on the GSS, the AI agents achieved an accuracy of around 69% against the initial human responses. When adjusted for human variability, the AI’s accuracy rose to approximately 85%. These figures exceeded the performance of simpler, demographic-based models by 14 percentage points, highlighting the sophistication of the AI’s ability to emulate individual behavior.
Park, quoted in New Scientist, expressed optimism about the potential applications of this technology, particularly in policy-making. “We can build an agent of a person that captures a lot of their complexities and idiosyncratic nature,” he said. By simulating human behavior with such precision, policymakers could potentially test the impact of proposed policies in a more nuanced and effective way.
However, the study’s implications have not gone without criticism. Experts interviewed in New Scientist raised concerns about the ethical dimensions of replicating individual personalities so accurately. Catherine Flick of Staffordshire University cautioned that AI agents, despite their accuracy, cannot truly grasp what it means to be part of a community, warning against over-reliance on simulated behavior.
Richard Whittle of the University of Salford noted the potential for AI-driven simulations to revolutionize areas like political messaging and market research but stressed the limitations of AI in adapting to significant changes in human contexts.
Park emphasized that the study was conducted under strict ethical guidelines, with participants able to withdraw their data and AI agent at any point. “This is a huge opportunity, but we are also extremely cautious about not setting a bad precedent,” he said in the New Scientist report. The team underscored their commitment to ensuring the technology empowers individuals rather than raising societal concerns about misuse or invasive applications.