YouGov says its test of AI “digital twins” found that adding detailed personal ratings improved how closely the simulations matched the people they represent. The company is challenging the research industry to adopt shared standards for validating these models, which are increasingly used to simulate public opinion.
Watch Desk analysis
What happened
In a report published on 10 November, YouGov says it created computational “E-twins” for 1,008 identity-verified panellists, then asked the real people and their simulated counterparts the same new questions. Sixty-nine per cent of panellists replied within 48 hours, giving YouGov answers to compare.
The company reports that adding each panellist’s ratings of more than 4,000 entities raised the share of answers within one point on a five-point scale from 78% to 90%. For fiction books, a category withheld from training, it says 90% of E-twin ratings were still within one point of the person’s response. YouGov also says agreement with open-ended answers varied by the twin’s confidence rating: 81% for high-confidence answers, 72% for medium and 67% for low.
Why it matters
If businesses use simulated audiences to make decisions, a result averaged across a group may hide whether an individual simulation actually resembles the person it stands in for. YouGov argues that comparisons with real people answering the same questions offer a stronger test than relying on ready-made datasets or aggregate patterns.
The findings come from YouGov’s own study of its own simulation approach. The company says exact agreement was 58% and that it has not yet measured how often panellists repeat their own answers, so it cannot say how close that result is to a ceiling. Those are useful limits, not reasons to skip the more basic question: what counts as good enough?
Our read
The valuable part is the challenge to show the working, not a claim that the digital-twin problem is solved. YouGov gives specific accuracy figures and a test against real respondents; the next step is comparable methods that let clients judge other providers on the same terms. Simulated people may be convenient, but convenience is not a validation standard.
What to watch
- Whether other simulation providers publish comparable tests against the people they model.
- YouGov’s promised measurement of how often panellists repeat their own answers.
- Whether the industry adopts shared validation standards, and what those standards require.
Discussion spark: Should AI-generated audience research be treated as useful evidence before it is validated against real people, or should that validation be a prerequisite for using it in business decisions?
Sources and evidence
- The YouGov challenge to all AI simulators (10 November 2026, 00:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.