YouGov (AIM:YOU) published a white paper on Wednesday validating the accuracy of its AI-generated "digital twins" against the real people they are modelled on, a test the London-listed research and data analytics group says exposes weak methodology used elsewhere in the industry.
The company built 1,008 LLM-based simulations, called "E-twins", of government ID-verified members of its proprietary panel, then put identical new questions to each twin and its human counterpart; 69% of panellists responded within 48 hours, giving a direct accuracy check rather than a comparison against public datasets.
Twins trained on panellists' ratings of more than 4,000 brands and entities matched their human within one point on a five-point scale 90% of the time, up from 78% when trained on profiling data alone.
The twins hit the same 90% match rate on fiction books, a topic excluded from their training data.
Panellists judged their twin's open-ended answers as describing their own views 75% of the time overall, rising to 81% for answers OriginalVoices, the AI simulation platform YouGov worked with, had flagged as high-confidence.
YouGov said much of the evidence simulators currently offer rests on public datasets that may already sit inside the training data of underlying language models, letting a twin appear to know an answer it has merely absorbed, with no way to recontact respondents to check.
"I challenge everybody in the industry to work together to establish standards for verifying and validating simulated data," said Stephan Shakespeare, chief executive and co-founder of YouGov, adding that trust in simulation "can only be achieved through widely agreed standards."
Read more about Yougov
- 13/10/26YouGov launches £15m buyback in place of dividend
- 06/08/26YouGov taps new CEO, plans buyback, extends debt facility
The company is calling on rivals to validate twins against individual humans, test them on unseen topics, publish methods and misses, and predict real-world outcomes in public.
YouGov flagged a gap for future work: the current study did not measure how often a person repeats their own answer to the same question, a ceiling against which a twin's accuracy must ultimately be judged, which it plans to test in the next iteration of the research.