Skip to content
News

YouGov tested 1,008 AI twins of real people. Here is what it found

On 10 November 2026, the research firm YouGov published a white paper that tests AI twins against the real people they were built from. The twins came close to their humans most of the time, and matched them exactly a little over half the time. YouGov also asked the rest of the industry to prove its own twins the same way.

What YouGov did

The paper is called "An evaluation of digital twins based on YouGov panel data". YouGov worked with OriginalVoices, an AI platform that specialises in simulation. Together they built 1,008 "E-twins", which are LLM-based simulations of the attitudes of real YouGov panelists whose identity had been checked against government ID.

The test was simple. The same new questions went to each twin and to the real person behind it. 69% of the real people answered within 48 hours, which let YouGov compare 693 pairs directly.

The main results

  • Close answers: when the twins were also trained on each panelist's ratings of more than 4,000 brands and other entities, they came within one point of the human on a five-point scale 90% of the time. Twins built on basic profile data alone reached 78%.
  • New topics: fiction books were left out of the training data. The twins still came within one point 90% of the time on them.
  • Exact answers: the twins gave exactly the same rating as their human 58% of the time.
  • Open questions: 75% of the time, panelists said their twin's written answer described their own views.
  • Confidence: each written answer was marked high, medium or low confidence. Panelists agreed with 81% of high, 72% of medium and 67% of low confidence answers.
  • Bias: YouGov reports no pattern of weaker results by age, gender, race or education.

What the numbers do not say

These are YouGov's own results about a product it sells, and we found no independent replication yet. YouGov itself says it has not measured how often people repeat their own answer when asked twice, so it cannot yet say how close 58% is to the best possible score. It plans to test that next and publish the result.

The paper also treats written answers as something to check. YouGov says open-ended output still needs quick validation by the real person, and that confidence levels show when a twin can be relied on and when you should talk to people instead.

The challenge to other providers

YouGov CEO Stephan Shakespeare used the paper to call for shared standards for checking simulated data. The company asks every provider of AI twins to:

  • validate each twin against the individual human it copies
  • test twins on topics they were never trained on
  • publish their methods
  • show where their twins miss
  • predict real-world outcomes in public

YouGov argues that tests on public datasets alone can hide twins that drift toward the average or toward stereotypes.

What it means if you use or build a twin

The study is about market research, but the lesson carries over to any digital twin of a person. A twin is most useful when the real person can check it.

  • Feed it real material. In the study, more of the person's own answers lifted accuracy from 78% to 90%.
  • Check it against the person it copies. Ask the same question to both and compare.
  • Watch the gaps. Low confidence answers were wrong more often, so a twin should say when it does not know.
  • Keep the human in charge of anything personal. Our guide on creating a digital clone of yourself covers consent and control.

How TalkMie handles this

A TalkMie twin answers from what its owner teaches it: text, voice notes, documents and photos. Questions it could not answer are collected in one list, so the owner knows what to teach it next. You can start a free twin or ask the founder's twin a question and judge the answers yourself.

Questions

How accurate were YouGov's AI twins?

Twins trained on each panelist's ratings came within one point of the real person on a five-point scale 90% of the time, and gave exactly the same rating 58% of the time. These are YouGov's own figures from its November 2026 white paper.

What is an E-twin?

It is YouGov's name for an LLM-based simulation of a real panelist's attitudes. The real person is called the H-twin.

Can an AI twin replace the real person?

The study does not claim that. YouGov says written answers still need quick checks by the real person, and that low confidence answers are less reliable.

Sources

Make a digital twin of yourself.

It answers in your words, speaks in your own voice and turns what it knows into short videos. Live in about five minutes, free to start.