An experiment simulating more than 1000 real people using the artificial intelligence model behind ChatGPT has successfully replicated their unique thoughts and personalities with high accuracy, sparking concerns about the ethics of mimicking individuals in this way.
at Stanford University in California and his colleagues wanted to use generative AI tools to model individuals as a way of forecasting the impact of policy changes. Historically, this has been attempted using more simplistic rule-based statistical models, with limited success.
“We really had to simplify human behaviour a lot to make these models,” says Park. “What we have the opportunity to do now is create models of individuals that are actually truly high-fidelity. We can build an agent of a person that captures a lot of their complexities and idiosyncratic nature.”
Advertisement
To build their AI-generated agents, the team recruited 1052 people in the US, representative of the wider population. Each participant took part in a 2-hour speaking interview with a voice-enabled version of GPT-4o, the most sophisticated version of the model that powers OpenAI’s ChatGPT tool. The AI was given a general script to follow but also instructed to listen to the answers in real time and adapt its questions as needed, asking participants for the story of their life and views on societal issues. The researchers then fed an AI-generated transcript of each interview into a different instance of GPT-4o and asked the model to imitate each person.
The team put each of these AI agents through a range of tests, including a long-running social attitudes survey called the General Social Survey (GSS), an assessment of the “big five” personality traits, five behavioural economic games and five social science experiments. The original human participants were also given the same tests twice, two weeks apart.
In general, the AI agents closely followed the participants’ responses to tests. The humans didn’t give identical answers two weeks later, with a match of around 81 per cent between the two testing sessions for the GSS. In turn, the AI agents had a raw accuracy of around 69 per cent against the first testing session, meaning they were essentially 85 per cent accurate when considering the differing human answers between the two sessions. The accuracy figures were similar, if slightly lower, for the other tests. The generative agents also outperformed simpler, demographic-based models at matching the responses of individuals by 14 percentage points.
“We didn’t quite know what we were expecting,” says Park, though he suspected that it would be possible for AI models to do well on these tests given anecdotes of ChatGPT users finding the AI could intuit facts about them. “It was a nice scientific validation.”
Park says his “core motivation” is to provide better tools for policy-makers to test out the impact of their proposals in a more nuanced way than broad-brush modelling of populations. While some may worry that the ability to replicate human emotions would be a gold mine for marketers, the ethical approval the researchers sought from participants limits the use of the agents and their underlying data “strictly for academic purposes”, he says.
“The potential of effective simulated human behaviour to trial the impact of policy is vast,” says at the University of Salford, UK. Politicians could also use these tools to test out messaging, he says, in a way that is quicker and more cost-effective than current methods like focus groups or polling. “The ability to trial campaign strategy on a simulated representative group would be invaluable.”
Yet Whittle does sound a note of caution. “Human behaviour is delightfully complex and context dependent,” he says. “Simulated groups may be very useful, but if the context changes significantly, human response will likely still be the go-to.”
Being too quick to rely on AI agents in this way could create issues, says at the University of Staffordshire, UK. “While I think this is really interesting research, I think that it has some quite problematic potential,” she says, “The [agents are] never going to understand [anything], because they can’t understand what it is to be in a community.”
The work also raises broader questions about the ethics of mimicking individual humans’ behaviour so accurately. While there is no suggestion that the AI agents are alive or conscious in a Black Mirror-style simulation, even Park and the team were keen to engage with such concerns.
“We take the ethics extremely seriously,” says Park. He points out that any participants who want to can withdraw their data – and therefore their AI agent – from the study or from being accessed by other researchers. “This is a huge opportunity, but we are also extremely cautious about not making this feel to set a bad precedent that’s going to impact our society in ways that, as you’re mentioning, are creepy or not empowering to individuals.”
arXiv
