Artificial intelligence can predict how you might answer a survey, but that does not mean it understands you, according to a new study published in Nature. Led by Harvard psychology researcher Ashwini Ashokkumar, the research found that GPT-4, a large language model (LLM), could forecast the outcomes of social science experiments with surprising accuracy, yet it systematically overestimated effect sizes and lacked genuine insight into human behaviour.
Study Methodology and Key Findings
The researchers assembled 70 real experiments conducted in the United States involving almost 120,000 participants. They provided GPT-4 with descriptions of hypothetical respondents, experimental messages, and survey questions, asking it to estimate responses under different conditions. Comparing the model's predictions with actual results revealed a strong correlation, and GPT-4 could often distinguish between more and less effective interventions.
However, the model systematically estimated effects to be about twice as large as the real results. "Useful forecasts are not the same as understanding," the study warns, echoing concerns raised by US scholars Lisa Messeri and Molly J. Crockett about AI creating "illusions of understanding."
Implications for Social Science Research
The findings suggest LLMs may capture meaningful patterns in text-based US survey experiments, but they are not a reliable shortcut around human research. The study notes that combining LLM predictions with human forecasts was more accurate than either source alone. "The most useful future may not be AI replacing human researchers or research subjects, but AI helping researchers decide where to direct scarce human resources," the authors state.
Researchers often run small pilot studies before expensive experiments, and LLM-generated forecasts could supplement these pilots. For example, simulating how different demographic profiles respond to vaccination messages or policy framings could refine interventions.
The Temptation and Risks of 'Silicon Sampling'
The concept of "synthetic respondents" or "silicon sampling" is gaining traction in polling, market research, and public consultation. Proponents see opportunities for faster, cheaper testing, but critics warn it could undermine trust if simulations are presented as real public opinion. "A synthetic sample draws on patterns encoded in the model's training data, prompt design and guardrails. It may reproduce elements of human judgement, but it lacks lived experience, local knowledge and a real stake in the issue being studied," the study explains.
This gap is especially important for emerging issues, marginalised communities, fast-moving events, and populations poorly represented in online data. The study found the model performed broadly well across demographic groups but identified some differences in accuracy favouring white and Republican samples in the US context.
Potential for Misuse and Need for Safeguards
The same predictive capability can be misused. The authors tested whether GPT-4 could identify social media content likely to reduce COVID vaccination intentions. While the model may refuse to generate anti-vaccination messages directly, it could still help identify which harmful messages from an existing set were likely to be most effective. "This highlights the need for safeguards that go beyond blocking obviously harmful prompts," the study warns.
Research using proprietary LLMs like GPT-4 is also vulnerable to changes by model providers, making it hard for other researchers to verify or repeat AI-based findings. The central lesson, according to the authors, is that LLMs may become valuable instruments for social science, but there is a risk they may get enough things right to make us think we understand more than we do.
