In August 2025, NewsGuard found that the ten most-used chatbots repeated false information 35% of the time when tested against false claims circulating at the time. A year earlier, the rate was 18%. In the same audit, the share of answers debunking the claims rose from 51% to 65%.
The change came as chatbots became less likely to refuse an answer. Their non-response rate fell from 31% to zero, meaning all ten systems answered every test prompt. NewsGuard’s combined failure measure, which counts either repeating a false claim or refusing to answer, fell from 49% in August 2024 to 35% in August 2025, even as direct repetition of falsehoods rose. Real-time internet searches exposed chatbots to information of varying reliability.
Election testing has exposed that weakness before. In 2023, AlgorithmWatch and AI Forensics systematically asked Bing Chat, Microsoft's assistant, about regional elections in Bavaria and Hesse. They also tested it on Switzerland's federal election. One-third of its answers contained factual errors, and in 40% of cases it avoided the question. The study covered three European electoral settings, reporting fabricated or misattributed political information as well as non-responses.
Some mistakes involved basic candidate information. Bing Chat attributed a scandal to the wrong Bavarian leader. It gave vote estimates that diverged from that day's polls, and listed a politician who had already withdrawn as a candidate.
A chatbot generates likely text from its training and information it finds online; it does not consult a verified database. A false claim on a prominent webpage can therefore appear in an answer written with the same polished confidence as a reliable report.
That confidence is especially risky when the subject changes quickly. Polls and vote estimates can become outdated. Attributed remarks and allegations of a late-breaking scandal also need checking. Postal-voting deadlines can change, and polling-place details warrant direct confirmation from official information.
Research published in Hipertext.net found another limitation. In a study of five language models answering questions about media and news cases in País Vasco, all favoured national sources over local and regional ones. Some gave readers little information they could trace. That gap matters for provincial candidates and debates on regional television, which may receive less attention than national politics.
The findings fit a wider problem in assessing technology claims. A separate RUSSPAIN report on European controls over Chinese vehicles and drones covered Spain's CNI assessment of risks near a possible SAIC plant in Ferrol; the earlier report concerned a different technology issue.
Usage is growing. The Reuters Institute's Digital News Report 2025 found that 7% of respondents worldwide had used an AI chatbot for news in the previous week. Among people under 25, the figure was 15%.
For stable topics such as how the D'Hondt system works, a chatbot can provide a starting explanation. For changing claims, readers can ask for sources and open them, then check whether each one supports the answer. A precise-looking figure is not proof of accuracy; important details should be compared with a news outlet that identifies its reporting and can correct errors.
The European AI Act establishes transparency obligations for generative AI systems. The source argues that companies developing them should explain how they handle election-related queries.