A recent Pew Research Center study asked digital twins of human panelists to reproduce nearly 300 survey responses. Synthetic estimates differed from human data by about 12 percentage points on average. The models also tended to flatten public opinion.
A more pointed warning comes from Peng et al. (2026). Their digital twins were given more than 500 responses from each person. Across 19 preregistered studies and 164 outcomes, twin-human correlations averaged only about .20, with problems including insufficient individuation, stereotyping and hyper-rationality.
Those findings should be taken seriously. Rich data alone do not create a reliable synthetic person. But there are reasons for optimism. Zhang et al. (2026) found that personal interview data improved the ability of LLMs to reproduce human responses, while Panfilova et al. (2026) found that longer interview transcripts yielded less error in modeling psychological measures. The issue isn’t whether synthetic respondents “work,” it’s how they are modeled.
At Breakthrough Research, we approach this less as persona simulation and more as respondent reconstruction. We build a deep model of each individual based on an extensive qualitative interview augmented by a robust set of quantitative questions. We then use a proprietary retrieval process where new questions are answered based on the most relevant evidence from the training model. The goal is to preserve the person’s distinctive decision signature, including preferences, motivations, contradictions and uncertainty, pulling them away from a generic prototype.
The response process is also important. Our approach first forms a substantive judgment from the respondent’s perspective and only afterward translates it into the survey’s response structure. We validate not just averages, but calibration, heterogeneity, variance, covariance and whether synthetic data produce the same strategic conclusions as human data.
We have also learned that no single synthetic approach fits every question. Stable preferences may work well with survey-style responses; context-dependent decisions may benefit from other approaches, such as behavioral simulation. Tailoring the approach to the question yields more accurate outcomes.
Pew and Peng et al. offer reasons for caution. But those weaknesses should lead to design requirements for better synthetic respondents, not reasons to stop building them.
