The team, led by Stanford computer science doctoral student Junseong Park, recruited 1,000 people diverse by age, gender, race, region, education, and political ideology. They received up to $100 for their participation. Through interviews with these individuals, the team created agent clones of those individuals. To test how well the agents imitated humans, participants took a series of personality tests, social surveys and logic games, each twice, two weeks apart. Agents then completed the same exercise. The results were 85% similar.
“I think that’s ultimately the future if you can have a lot of little ‘you’s running around and actually making the decisions that you make,” Joon says.
In the paper, the clones are called simulation agents, and the impetus for creating them is to make it easier for researchers in the social sciences and other fields to conduct costly, impractical or unethical research on real human subjects. If we can create AI models that behave like real people, we can use them to test everything from how well interventions on social media combat misinformation to behavior that causes traffic jams.
These simulated agents are slightly different from the agents that dominate the work of leading AI companies today. These models, called tool-based agents, are models that are built to perform tasks for users rather than converse with them. For example, you might be able to enter data, retrieve information you have stored somewhere, or one day even book a trip and make an appointment. Salesforce announced its own tool-based agents in September, Anthropic in October, and OpenAI says it plans to launch some in January. bloomberg.
Although both types of agents are different, they share commonalities. Research on simulated agents like the one in this paper has the potential to lead to more powerful AI agents overall, says John Horton, an associate professor of information technology at the MIT Sloan School of Management who founded a company that conducts research using AI. Simulation participants.
“This paper shows how you can do a kind of hybrid: you can use real humans to create personas and then use them programmatically/in simulations in ways you can’t do with real humans,” he said. MIT Technology Review By email.
There are caveats to this study, the most important of which are the risks it points out. Just as image generation technology has made it easy to create harmful deepfakes without people’s consent, any agent creation technology will make it easier for people to build tools to anthropomorphize others online and say or approve things they didn’t intend. It raises questions about. .
The evaluation method the team used to test how well the AI agent replicated its human counterpart was also fairly basic. This included a general social survey that collected information about a person’s demographics, happiness, behavior, etc., and an assessment of the Big Five personality traits (openness to experience, conscientiousness, extraversion, agreeableness, and neuroticism). Although these tests are commonly used in social science research, they do not pretend to capture every unique detail that makes up who we are. AI agents also did worse at replicating humans in behavioral tests such as the “Dictator Game,” which is meant to shed light on how participants consider values like fairness.
