TL;DR

British politicians are trying out policy messages on AI personas built from real polling data, Bloomberg reported on 22 October in a piece syndicated by The Japan Times. AI minister Kanishka Narayan says synthetic focus groups were among the first things he used as an MP, and former Downing Street chief secretary Darren Jones built one in Anthropic’s Claude. Critics, and research from Pew, question how far the results can be trusted.

Who is using it

Startups such as Artificial Societies and Electric Twin sell the service, as do bigger, established players Toluna and GWI. At the cheap end, Synthetic Respondents, a London firm, asks $16 (£12) a project, a price it puts at one-thousandth of traditional research.

Speaking at Labour’s conference last month, at a fringe meeting run by Campaign Lab, Narayan said that “even as an individual backbench MP” he could now gauge public opinion far more cheaply than before. He has used them to test the reasoning of speeches in draft and to aim messages at voters in his Vale of Glamorgan seat, and Labour has used them to train volunteers in doorstep conversations. Jones said he uses his Claude-built version “to guess sentiment on an issue” before monthly town halls. Narayan’s wider case for state AI capability is in his BBC interview this week.

The accuracy question

Pew’s study, published last month, built AI “digital twins” of US panel members and found that across nearly 300 questions the AI answers were, on average, 12 points away from the humans’, and they struggled with emerging topics like data centres. Electric Twin says its audiences come within roughly 4.5 points of real responses, though on a different, aggregate measure.

Chi Onwurah, chair of Parliament’s science and technology committee, said she would need to understand how a tool was trained and weighted before relying on it. Luke Tryl, whose organisation More in Common runs the human kind, said the point of a focus group is to “bring out the idiosyncrasies and nuances”. A June paper by Dario Krpan and colleagues at LSE cautioned that politicians could begin treating synthetic views “as if they were the views of actual people”.

Looking forward

Electric Twin co-founder Ben Warner, who worked in Downing Street during the Covid response, sees the main value at the start of policy design, when a fast read is what is needed. In our view, that is the defensible use: a cheap first pass for backbenchers who could never afford commissioned research. The risk the critics describe is the step after, when agents that “do not vote”, as Yale’s Claudio Novelli put it, start standing in for people who do.