Research
Synthetic twins, tested against real people. Now we’re matching minds, not just answers.
The method is peer-reviewed. Across 373 real people, three products and four AI models, twins’ answers tracked their real counterparts closely in aggregate (correlations of .81 to .91). The paper is in press at Behavioural Public Policy (Cambridge University Press), grounded in behavioural science from the UK’s #1 ranked university in Psychology, the London School of Economics and Political Science.
Paper I
In pressCan AI represent the way real people think and decide?
We compared Dipio’s synthetic twins against 373 real people, across three products and four AI models. In aggregate, their answers tracked real humans closely. But the paper shows the limit too.
Matching what people say is not the same as matching how they think, and closing that gap is the mission we’re on now.
From matching outputs to matching minds.
The field judges a synthetic user by one thin test: does its answer resemble a human’s? That says nothing about why it answered, or whether it would answer the same way tomorrow. We’re building the layer underneath: twins that reason from the same psychological mechanisms as real people, their traits, values, heuristics, and biases, so the resemblance holds when the question, the model, or the task changes.
Matching outputs is a benchmark. Matching minds is the science.
Twins Studio, the product built on this research, is in early access. Every interview you run leaves a twin of that person, a living model you can keep asking follow-up questions long after the call ends. See how twins work