One third of information provided by chatbots during the election was incorrect, finds cross-party think tank Demos

Text-based AI services such as chatbots and AI-powered search results routinely provided inaccurate information about elections ahead of this month’s polls, according to new evidence from the cross-party think tank Demos.

Errors included getting the date of election day wrong, giving wrong information about the need for voters to bring ID, “hallucinating” a candidate, and making up an expenses scandal on one occasion, and a nepotism scandal on another.

The think tank tested how five popular AI services performed on a single day during the 2026 Scottish pre-election window and found that one third (34.1%) of responses across chatbots contained factual errors, whilst reliability varied significantly across services.

The five services tested were ChatGPT, Google Gemini, Google AI Overviews, Grok and Replika – which are all owned by US corporations.

Researchers utilised a red-teaming methodology to prompt AI services with 75 questions about the Scottish elections, testing across three anonymised Scottish constituencies.

The findings provide a live snapshot of the threat generative AI poses to democracy across the UK, whilst new legislation struggles to keep pace with the technology.

As part of the research Demos also carried out a nationally representative survey which suggests these findings have implications across the UK, as 1 in 5 adults (20%) report using an AI chatbot or AI search service to find information about the local and devolved elections this month.

Key findings:

• Persistent unreliability across the board: One third (34.1%) of responses generated by text-based AI services contained inaccurate information. Examples of errors included: getting the date of the election wrong, hallucinating candidates, incorrect advice on voting procedures, and made-up political scandals. The results suggest AI chatbots and related services have serious shortcomings as information sources during elections.
• Error rates varied significantly across services: Replika performed the worst: 56.4% of its responses contained errors. ChatGPT – widely considered the most popular service for UK users – also had serious issues with accuracy, providing responses that contained errors 46.2% of the time. While Google Gemini and Grok fared significantly better, their results still included inaccuracies with a 21.8% and 8.97% error rate respectively. This demonstrates that across the most commonly used AI chatbots, none can be relied upon for complete accuracy.
• Accuracy problems were demonstrated for all constituencies tested: The rate of errors in responses was relatively consistent across all three constituencies: 37.9% in Constituency A, 30.3% in Constituency B, and 29.7% in Constituency C. This suggests that areas where high-profile politicians are based may be most prone to inaccuracies.
• Sources were inconsistent and unreliable: Half (46.9%) of chatbot responses did not come with citations or links to back their claims and two thirds (66.3%) of responses did not cite an official source of information on the Scottish elections, such as the Electoral Commission or Scottish government. These gaps undermine users’ ability to evaluate accuracy or levels of political bias.
• Services offer advice on tactical voting that could affect the result: Researchers tested whether services would offer advice if asked who to vote for to prevent a specific named party from getting into power. All services besides Google AI Overview provided such advice on tactical voting every time they were asked (12 of 12 responses, excluding AI Overview). Google did not generate an AI Overview at all for this question. While some services did not outright select a preferred candidate (e.g. ChatGPT, Gemini), others did so either explicitly (Replika) or implicitly (Grok).

The threat of a national-scale misinformation crisis

Polling carried out by Demos between the 30th April and the 6th May 2026 suggests the inaccuracies identified in chatbot responses could have severe implications for democracy across the UK.

• Use is widespread and growing: 1 in 5 adults (20%) report using an AI chatbot or AI search service to find information about the local and devolved elections in the run up to May 7th – that’s equivalent to over 10 million UK adults.
• The public are concerned about accuracy: Half (47%) are worried about AI chatbots sharing inaccurate information about elections and candidates.
• Public trust is very low: Half 49% do not trust AI chatbots for election-related information – and the services are as distrusted as social media (also 49% distrust).

Policy recommendations

Demos has proposed a series of short and long term recommendations to begin to tackle the challenges identified in this research. This is in the absence of any current regulation on AI, which remains highly unpopular with the public. According to Demos’ polling, the government’s deregulatory approach to AI is unpopular: Half (52%) oppose an unregulated approach when it comes to AI and elections.

• Make existing UK law LLM-ready: Clarify how election law and defamation law apply to AI-generated claims, and review how to ensure key election protections can be updated to address risks from LLMs. Use the Representation of the People Bill to bring in protections, before the window for amendments passes.
• Mandate minimum AI election safeguards: Legislate to establish a baseline requirement for election safeguards that AI providers must implement at all times, with heightened requirements during pre-election windows.
• Ensure transparency and data access: Require text-based AI services to provide independent researchers with access to internal data, training sets, and live data during election windows to ensure public-interest accountability.
• Invest in trust by supporting AI text detection technologies: Equip the public to distinguish AI outputs from human-made text by investing in the development of textual watermarking and AI detection. Fund innovation in AI text detection and support the development of new cross-industry standards.

Azzurra Moores, Associate Director of Information Ecosystems said:

“Our research shows obvious and substantial inaccuracies in the information generated by AI chatbots about The Scottish election.

The experiment gives us a live example of what we mean by ‘unreliable information’ being given to the public in a regional context – but we’re very clear that the issue is not unique to Scotland, or the constituencies we tested in. This is a UK-wide, if not global, concern.

The accessibility of these AI-tools – which are all developed and run by US corporations – is widespread in the UK, but 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.

Building these protections is nothing short of an urgent requirement, because a healthy democracy relies on trustworthy information. Our recommendations touch on a range of legislative action that could be taken, but the fastest and most effective route is through the Representation of the People Bill.

AI chatbots are a rapidly growing source of influence over our democracy – the government should not miss this opportunity to put the right safeguards in place.”

// ENDS //

NOTES TO EDITORS

Methodology: Red-teaming exercise

This study utilised an established AI red-teaming methodology: two analysts prompted five AI services with 75 questions about the Scottish elections over one day during the pre-election window. In total, 375 questions were asked across 15 separate conversations with the five services.

The five AI services covered were ChatGPT, Google Gemini, Google AI Overviews, Grok, and Replika, a companion chatbot.

The study tested these across three Scottish constituencies: Constituency A (a constituency in a large city where a prominent politician is running), Constituency B (a hotly contested constituency in a large city with redrawn boundaries) and Constituency C (a rural constituency).

The names of the specific constituencies or candidates are not disclosed in order to protect the identities of those contesting the races and to avoid placing undue attention on individual candidates.

The testing was designed to replicate the experience of an average Scottish user. To do so, a Virtual Private Network (VPN) was used to appear as a Scottish internet user and only using the free tier of each service to best mirror a realistic user for this case.

Data was collected on one date to maintain consistency across the services and ensure that fact-checking was as rigorous as possible.

Measures were taken to ensure that each conversation about each constituency counted as an independent test and was not biased by other conversations.

Methodology: Polling

Polling conducted by Opinium on behalf of Demos.

Fieldwork conducted 30th April – 6th May 2026.
Nationally representative sample of 2,005 UK adults aged 18+.
Full polling tables available on request.