AI Chatbots Spread Election Misinformation 34% of the Time
Key insights
- Replika had a 56% error rate on election questions; ChatGPT reached 46%, while Grok performed best at 9%.
- Errors included invented candidate scandals and false voter ID requirements, not just vague or outdated information.
- One in five UK adults used AI chatbots or AI-powered search for local election information during the study window.
Why this matters
AI products are now functioning as de facto civic infrastructure for millions of voters, yet none of the major platforms tested met a basic accuracy threshold for election-critical facts. A 34% error rate at that scale of adoption means misinformation is being delivered with the confidence and UX of authoritative search, making it harder to correct than a social media post. The Electoral Commission push for legal duties on AI platforms signals that election-law liability for AI outputs is no longer hypothetical and could reshape how frontier labs design and restrict model behavior around civic topics.
Summary
A Demos study caught AI chatbots fabricating candidate scandals, inventing wrong election dates, and falsely telling Scottish voters they needed photo ID to vote. Across 75 questions about three real Scottish constituencies tested during the pre-election window, 34.1% of responses contained factual errors, with some tools performing far worse than others.
Replika led the failure rate at 56%, followed by ChatGPT at 46%. Grok came in lowest at 9%. The platforms tested included ChatGPT, Google Gemini, AI Overviews, Grok, and Replika.
Essentially: (OpenAI, Google, xAI, Replika) are being used as civic information sources by a population that largely trusts them.
- 1 in 5 UK adults used AI chatbots or AI search to look up local election information.
- Errors weren't just omissions: the study documented invented scandals about real named candidates.
- The Electoral Commission is now pressing UK ministers for new legal duties on AI chatbot platforms.
The photo ID error is particularly pointed: that claim directly mirrors a real ongoing UK voter ID debate, meaning AI tools are injecting false certainty into an already contested policy space.
Potential risks and opportunities
Risks
- OpenAI and Google face direct regulatory exposure in the UK if the Electoral Commission's push for legal duties on AI platforms advances through Parliament before the next scheduled national election.
- Replika's 56% error rate on civic questions, if widely reported, could accelerate platform bans or use restrictions in EU member states already scrutinizing companion AI apps under the AI Act.
- Candidates named in AI-fabricated scandals have actionable defamation claims in UK law, and a single documented case could set precedent forcing chatbot providers to implement pre-election content restrictions.
Opportunities
- Civic-tech and election-integrity nonprofits (Full Fact, Democracy Works) are well-positioned to formalize AI audit partnerships with electoral commissions across the UK and EU ahead of 2027 election cycles.
- RAG-based political information products with verified electoral data sources (candidate registers, official polling rules) have a clear market opening as platforms scramble to reduce liability.
- UK legal firms specializing in defamation and electoral law gain a new advisory product line helping candidates and parties document and respond to AI-generated false claims about named individuals.
What we don't know yet
- Whether any of the tested platforms (OpenAI, Google, xAI) were notified of specific errors before publication and whether corrections were deployed before the Scottish election.
- How Grok's 9% error rate was achieved relative to the others: whether it reflects retrieval architecture, knowledge cutoffs, or refusal behavior on uncertain queries.
- Whether the Electoral Commission's push for new legal duties includes liability for AI Overviews specifically, given Google's integrated placement in organic search results.
Originally reported by theguardian.com
Read the original article →Original headline: Demos Study: AI Chatbots Gave Voters Wrong Information in 34% of Scottish Election Questions — Invented Scandals, Wrong Dates, False Voter-ID Claims