MANY MINDEDAI Behavioral Twins Show Human Oversight Still Critical
A study led by Columbia Business School’s Olivier Toubia created digital twins from 2,000 U.S. participants responding to 500+ questions in a 2025 research project. These models achieved above-chance accuracy in 19 behavioral experiments with an average error rate of 25%. The digital twins captured greater response variation than LLMs using only demographic data and demonstrated higher rationality than human participants. However, they exhibited homogeneity in responses, showed demographic bias (increased accuracy for more affluent and educated participants), and expressed higher trust in others while reducing concern about technological threats compared to humans.
The research—published in *Science Advances* on September 2, 2026—used open-source data from the *Marketing Science* journal. Key caveats include the study’s static question sets without dynamic interaction methods, accuracy trends observed only among participants with higher education and income levels, and digital twins performing similarly to demographic-only LLMs in overall accuracy. The dataset was downloaded 25,000 times, indicating interest but not real-world implementation.
This finding matters for the knowledge sector’s friction: it confirms AI agents remain incapable of replacing human researchers in complex behavioral contexts without oversight. For basic needs like equitable healthcare resource allocation (a concern noted by Penn State’s Hadi Hosseini), this means human judgment remains essential to avoid biased outcomes. What to watch: whether dynamic interaction methods can bridge the gap between static models and real-world behavioral complexity. The current limitations highlight that free, high-quality behavioral research requires human oversight—preventing AI from inadvertently worsening inequities in critical areas.
Source: Science News