Demos finds AI chatbots gave voters wrong information in 34% of Scottish-election questions

TL;DR:

• A Demos study, Electoral Hallucinations, found AI chatbots gave voters wrong information in 34% of 75 test questions about three real Scottish constituencies — including inventing scandals, wrong election dates and false claims that voters needed ID at polling stations.
• Companion polling found 20% of British adults had used AI tools to get information about the Scotland, Wales and English local elections — equivalent to 10 million UK voters — and Replika performed worst with 56% errors, ChatGPT 46%, Gemini 22%, and Grok 9%.
• Electoral Commission chief executive Vijay Rangarajan said the current legal framework “should go further” and is pressing ministers for clearer duties on AI platforms and stronger Ofcom enforcement powers.

The findings move the UK election-misinformation conversation from abstract concern to specific numbers. Demos’s tests were not designed to trip up the systems with adversarial prompting; they used 75 ordinary questions about three real-life constituencies. A 34% misinformation rate across the major free AI tools is not an edge-case finding. The Electoral Commission’s response — pressing for new legal controls — comes from a regulator that had already discovered half of voters in the 2024 general election saw misleading information.

A UK-specific test that highlights a UK-specific regulatory gap

The Online Safety Act, the Defending Democracy Taskforce and the Representation of the People Act collectively do not impose explicit duties on AI platforms to protect voters from misinformation during election periods. Demos associate director Azzurra Moores told the Guardian: “We don’t yet have the legislative framework to protect the public from misinformation, or our democracy from the knock-on impact of its circulation.” Demos proposes three specific moves: making AI companies liable under UK defamation and electoral law, mandating accuracy safeguards, and forcing AI firms to allow researcher access to their data and training sets.

The per-tool breakdown is informative because the same companies are deeply embedded across UK government, finance and public-sector AI rollouts. ChatGPT’s 46% error rate covers making up an expenses scandal, getting the election date wrong by two months, and inaccurate replies on voter eligibility. Google Gemini was wrong about a candidate’s position on assisted dying and claimed a police investigation involving the SNP was continuing. In nearly half of all responses across all tools, no official sources or external links were provided to support answers; where citations were given by ChatGPT, they were at least a year out of date 44% of the time. The Department for Science, Innovation and Technology said defending elections was “an absolute priority” but did not commit to amending the Representation of the People Bill. OpenAI told Demos the methodology was “not typically how ChatGPT uses its services” and might have used an out-of-date version.

Looking forward

The 2026 conversation now has a UK-grounded evidence base. Expect the Electoral Commission to formalise its proposed legal duties before the next devolved-election cycle, Ofcom to be asked to clarify which AI behaviours fall under existing Online Safety Act provisions, and DSIT to face renewed pressure on the AI Bill timetable. For UK businesses deploying chatbots for customer service or research support, the Demos methodology — accuracy-tested across 75 ordinary questions — is a useful benchmark to apply internally before any tool reaches a public-facing context.