U.K. politicians are testing their messages on AI-generated focus groups

U.K. politicians looking to gauge how British voters will react to their policies have a new cohort on which to test their ideas: AI-generated focus groups.

Startups Electric Twin and Artificial Societies, and more established research firms GWI and Toluna are using data from real focus groups and surveys to build AI personas able to react to political messages within minutes. Another firm, London-based Synthetic Respondents, charges $16 (£12) a project, which it says is a thousandth of the cost of a traditional piece of focus group research.

Critics warn that reliance on AI-generated focus groups could detach politicians from reality. But the technology also has the potential to transform their ability to sense-test their ideas, making it quicker and more affordable to gauge how real-life voters may react to policies, and giving backbenchers access to analysis that might previously have been outside their budget. One early adopter was Kanishka Narayan, elected to Parliament in 2024, and now Britain’s AI minister.

“The first thing that I remember I did as an MP was start to use synthetic focus grouping in thinking about what messages to deliver,” Narayan told a Campaign Lab fringe event at Labour’s annual convention last month. “You can now, even as an individual backbench MP, start to understand what people are thinking in a much cheaper way than ever before possible.”

The technology is broadening access to granular data that was “previously under the reserve of a central team in headquarters,” Narayan said. He added that he’d used the tool to challenge his own arguments in draft speeches and to help better target communications in his Welsh constituency, the Vale of Glamorgan. Labour has also used it to help train campaign volunteers with practice doorstep conversations.

Some MPs have taken the approach further. Darren Jones, former chief secretary to the U.K. prime minister, said he built his own AI focus group in Anthropic’s Claude using polling, election results and demographic data. “I mainly use it to guess sentiment on an issue before my monthly in-person town halls,” he said.

Companies are also getting in on the act. Artificial Societies said a Fortune 500 firm, which it declined to name citing confidentiality, used it to simulate the most likely question from each analyst on an earnings call. The Times named its business podcast ‘The Business’ based on Electric Twin research, while Octopus Energy’s research arm recently cited synthetic polling on support for data centers in a report.

There are drawbacks. Synthetic voters struggle with the cases where they would be most useful: minority viewpoints and new issues poorly represented in the training data. Critics contend they could also weaken the link between politicians and the people they represent, lead to the proliferation of AI-written speeches and discourage politicians from exercising independent judgment.

“Ultimately, these agents do not vote,” said Claudio Novelli of Yale University.

A Pew Research Center study published last month concluded AI models “are not an adequate replacement for traditional polling on topics of broad public importance.”

Pew built AI “digital twins” of U.S. panelists using their demographics and political views. Across nearly 300 questions, the AI results differed from humans by 12 percentage points on average, struggling especially with stereotypes and new issues such as data centers. There was also a tendency for digital voters to congregate toward the same answers more than their human counterparts.

“The whole point of focus groups is not to flatten but to bring out the idiosyncrasies and nuances,” said Luke Tryl of More in Common, which runs real focus groups.

Chi Onwurah, who chairs Parliament’s science and technology committee, said the tools would need to be accurate before she would use them.

“As with any focus group you have to understand the methodology behind its formation to use it effectively,” she said in an interview. “Specifically for AI that would involve a complete understanding of how it was trained, what guide rails were in place, weightings etc.”

Pew’s research highlighted problems that Ben Warner and Alex Cooper had to overcome when they co-founded Electric Twin in 2023. The company says its synthetic audiences come within about 4.5 percentage points of real responses, albeit using a different test from Pew’s, looking at aggregate responses.

The two men met while working in Downing Street on the COVID-19 response, and set up their new venture because they wanted policymakers to have more information about how people will respond to policies than they had.

Synthetic voters have a “really valuable role to play” early in a process, when a quick read is needed, Warner said. “Imagine the Treasury before their budget: they can’t really be like, ‘If we did this market moving thing, would you be pro it or against it?’”

Pollsters are also experimenting, with Ipsos following five AI personas of undecided voters through the 2024 general election, and Gallup testing AI agents built from in-depth interviews with about 1,000 of its panel members since May.

One risk academics see is the potential for the blurring of reality. Dario Krpan and colleagues at the London School of Economics warned in a paper in June that politicians may “begin treating the views generated by synthetic participants as if they were the views of actual people.”

Earlier this year, Axios updated an article to clarify that polling it cited was generated by AI firm Aaru rather than real people. And in August, a Los Angeles mayoral poll received a lot of attention before its publisher, Median Strategies said it had fabricated the survey as a social experiment in how purported polling can spread without verification.

Krpan and his colleagues said there’s little doubt synthetic polling could improve government, noting that policies have failed because they weren’t tested, and good ones have been overlooked because politicians wrongly assumed they’d be unpopular. But they also warned that it’s hard to replicate the back and forth of a focus group that can help tease out the reasoning of voters.

“There is an important distinction between predicting what people think and understanding why they think it,” they wrote.

That’s a sentiment echoed by Jason Mander, Chief Insight Officer at GWI, which this month is launching its own synthetic product designed for use by companies rather than politicians.

“It’s never going to be the same as going and speaking to a representative sample and actually hearing from them,” he said. “Humans are messy, they’re complex, they’re confusing, they contradict themselves.”