The UK Electoral Commission has called for new legal controls over misinformation from AI chatbots after a study revealed significant failures during the recent Scottish election. The investigation by thinktank Demos found that AI services provided misinformation in 34% of answers to election-related questions, highlighting the risks of unregulated AI platforms.
In its report titled "Electoral Hallucinations," Demos detailed a simulation involving 75 questions posed to five free AI tools, including ChatGPT, Google Gemini, and Replika. The research focused on three real-world constituencies to evaluate the accuracy of AI responses. The findings were alarming: tools invented fictitious scandals, provided incorrect election dates, wrongly claimed that Scottish voters needed ID at polling stations, and misassigned candidates to contests.
A commissioned poll of 2,005 British adults suggests the scale of the issue is significant. Roughly 20% of voters—equivalent to 10 million people across the UK—reported using AI chatbots or search tools to gather information about local and parliamentary elections in Scotland and Wales.
Vijay Rangarajan, chief executive of the Electoral Commission, is pressing ministers for legislation to hold AI companies accountable. He emphasized that voters require accurate information for democratic engagement and voiced concern that AI has made the spread of false information faster and more accessible. He advocated for clear duties on AI platforms to prevent algorithm-driven misinformation, which would provide Ofcom with the necessary powers for enforcement.
Azzurra Moores, associate director at Demos, noted that while these tools are widespread in the UK, they are predominantly operated by US-based corporations without a domestic legislative framework to mitigate their impact. She recommended making AI firms liable under defamation and electoral laws, implementing mandatory accuracy safeguards, and requiring transparency for independent data testing. According to the study, the companion bot Replika performed the worst, with a 56% error rate, including the creation of entirely fabricated candidates and corruption allegations.
