AI chatbots gave misleading voting advice ahead of Senedd poll

TL;DR:

• BBC Wales asked ChatGPT, Copilot, Gemini, Claude, Meta AI and Grok how to vote in Thursday’s Senedd election using six fictional voter profiles built with the National Centre for Social Research.
• Several chatbots returned inaccurate constituency data, listed candidates not on the ballot, named a deceased Senedd member among possible candidates, and gave conflicting party recommendations to identical voter profiles.
• Resultsense view: the BBC’s findings land beside Peec AI’s separate UK-wide research showing Reform UK is dramatically over-represented in LLM responses on British politics — together they signal a real democratic-process problem heading into the May local and devolved elections.

The investigation, published two days before polling day, fed each chatbot the same scenarios drawn from undercover voter personas. Floating voter “Lauren” — a Welsh HGV driver concerned about cost of living and the NHS — received conflicting advice: ChatGPT pointed her to Labour or Plaid Cymru, Grok to Reform UK. The chatbots’ descriptions of devolved-versus-reserved policy issues were broadly accurate; the candidate-level detail was where errors clustered.

Specific failure modes

Examples cited by the BBC include: Copilot returning the wrong constituency for a stated town; ChatGPT and Meta AI providing candidate names not actually on the lists; Gemini surfacing Hefin David, the former Labour Senedd member who died in 2025, as a possible candidate for Blaenau Gwent Caerffili Rhymni; Claude wrongly suggesting that Plaid Cymru leader Rhun ap Iorwerth had stepped down. Several chatbots also provided incomplete candidate lists, omitting names altogether.

OpenAI, Google and Microsoft all responded to the BBC saying their tools encourage users to verify information; the firms behind Claude, Meta AI and Grok were also asked for comment. AISI research cited in the report estimated that around 13% of UK voters used conversational AI for political information ahead of the 2024 general election — a baseline likely to have risen in the year since.

UK pattern, not a Welsh one

The Senedd findings sit alongside Peec AI’s separate UK-wide study, also published this week, which found Reform UK appears in 88% of Google AI Overviews on British politics and that Nigel Farage is referenced more than Keir Starmer across major models. The Welsh case is the local manifestation of a national pattern: AI systems trained on broad web content and now leaning on social media for current events are producing politically uneven outputs.

Looking forward

The Senedd, Scottish Parliament and English local elections all run alongside heightened LLM use. Voter-information accuracy is fast becoming an emerging-issue area for the Electoral Commission, the ICO and Ofcom — none of whom currently regulate LLM political content directly. UK firms running election-period communications should expect tighter scrutiny of any AI-assisted voter messaging in the next 12 months.