Expert: Dr Eoghan Cunningham, Lead author, Auditing AI Chatbots During the Galway West and Dublin Central Byelections (UCD Connected Politics Lab / University of Strathclyde), postdoctoral researcher at University College Dublin **What the researchers did.** In the weeks before the Dublin Central and Galway West byelections of 22 May 2026, a team from University College Dublin's Connected Politics Lab and the University of Strathclyde simulated ordinary voter behaviour: the same 194 election questions - from 'where do I go to vote?' and 'who is running?' to policy comparisons and questions about every individual candidate - were put to Anthropic's Claude, OpenAI's ChatGPT, Google's Gemini and xAI's Grok, on two separate occasions 14 days and 7 days before polling day (study authors: Eoghan Cunningham, James Cross, Louise Luxton, Timea Balogh and Derek Greene). **Where the answers failed.** The eleven basic factual questions were mostly answered well, because official guidance from the Electoral Commission, citizensinformation.ie and the returning officers is easy to find and the chatbots cite it. The failures came from retrieval and interpretation: ChatGPT left at least six names off the ballot when reporting who was running in the two constituencies, and Gemini reported that Gerry Hutch had won the fourth seat in the 2024 national election - he ranked fourth on first preference votes but did not win a seat. The authors' summary of the mechanism is that when a chatbot fails to retrieve the correct information while searching the web, it may still answer confidently, misinforming voters about their choices without signalling any uncertainty. **The sourcing split.** Wikipedia was the most frequently cited single source for almost every chatbot except Claude, whose top source was RTE. ChatGPT and Gemini never cited RTE at all, and the conservative Irish news site Gript.ie was Gemini's number one news source. In the Irish Times' breakdown of cited links, Claude leaned on RTE (20 per cent), ChatGPT on The Irish Times (17 per cent) and on Wikipedia (16 per cent), and Gemini on Gript.ie (7 per cent); Grok was the heaviest social-media user, citing Facebook for 6 per cent and Instagram for 5 per cent of its cited links. Because Wikipedia is crowdsourced, the researchers note that bad actors could edit entries about rival candidates and have that reflected in AI answers. **Who gets seen.** All four systems concentrated on the same handful of candidates and largely ignored those further down the field. Within that group, the choice of chatbot changed the picture: in Galway West, Míde Nic Fhionnlaoich came across as a middle-of-the-pack candidate on one system and as the most frequently encountered candidate on another. **Why the verdict is 'incorrect'.** A chatbot that omits six candidates from a ballot, or awards a seat to a candidate who never won one, is not just imprecise on a detail - it changes what a voter knows about who is standing and who won. The article's own framing from the study's director, Dr James Cross of UCD, is that most answers were good, 'but when it goes wrong, it goes really wrong'. The study's lead author, Dr Eoghan Cunningham, adds that the system is 'always guessing' on material like an Irish byelection. The findings point at a fix rather than a complaint: show which sources were retrieved and cited, keep unverifiable sources such as crowdsourced encyclopaedias and social posts out of the top of the citation list, and have the system state when it could not retrieve the record instead of answering without it. Context for scale: the Reuters Digital News Report found 7 per cent of audiences across 48 countries use AI to get news weekly, rising to one in six people under 25. Original study blog post: https://www.ucd.ie/connected_politics/blog/auditingaichatbotsduringthegalwaywestanddublincentralbyelections/ Source: https://www.irishtimes.com/politics/2026/05/21/are-you-using-ai-to-inform-yourself-on-the-byelection-researchers-warn-of-chatbot-errors/