AI chatbots got election information wrong 90% of the time in new study, including ChatGPT rivals

New research found that leading AI chatbots frequently provided incorrect election information
Researchers say systems like ChatGPT, Gemini, Claude, and Grok still struggle with sourcing and policy accuracy.
AI chatbots are becoming trusted information tools even before they are trusted around elections.

Artificial intelligence companies are interested in making their chatbots a source of information. But a new study suggests that election data remains a place where that trust may be outweighing technology readiness.

NewsBench, a project created by Forum AI that studies how AI systems handle journalism and news reporting, found that top AI chatbots repeatedly ran into problems when asked election-related questions. The findings add to growing evidence that conversational AI systems remain unreliable in one of the highest possible risk categories: helping people understand democracy itself.

“Ask a leading AI chatbot a question about the upcoming midterm elections, and there’s a 90% chance the answer will be wrong in some material way: a factual error, a clear partisan bias, a quote from a state-controlled foreign media outlet, or some combination of the three,” Forum wrote in a summary of the study.

Trust issues

AI systems frequently fail because they surface weak sources, incomplete information, or incorrect material before beginning to generate a response. The researchers found that recovery failures caused more than 70% of the errors observed. When systems successfully retrieved reliable information, they often responded correctly. Getting the right information consistently remained the most difficult challenge.

That problem becomes especially uncomfortable during elections. The chatbot seems confident regardless of its accuracy. Answers are written clearly with citations and authoritative language. Even incorrect information can appear trustworthy when presented with sufficient confidence.

Election-related studies increasingly show how dangerous that combination can be. Chatbots often combine precise details with inaccuracies in ways that are easy for users to perceive. The result bears no resemblance to the misinformation websites of earlier eras of the Internet. It looks like experience.

That distinction is important because people increasingly treat chatbots less as experimental software and more as infrastructure. OpenAI, Google, Anthropic and other developers routinely encourage users to independently verify important information. Companies continue to sign licensing agreements with publishers, partly in hopes of improving the quality of sourcing and reducing factual errors.

Electoral pressure

Even the most capable models ran into problems when source selection failed. Accuracy often decreased further when questions contained subtle inaccuracies or misleading assumptions, similar to how real users often ask questions online.

The moment creates additional pressure. AI companies are shipping increasingly sophisticated reporting tools, while regulators continue to move unevenly across different countries. Europe has put more pressure on transparency requirements. Other governments remain at an earlier stage in the process. Meanwhile, adoption continues to grow.

The long-term answer may involve stronger source attribution, more transparent retrieval systems, better provenance technology, and a more robust editorial infrastructure behind AI products. The challenge is that the elections do not wait for technology companies to finish improving their systems. Voters use the tools available today and it is clear that they need to be improved.

Google logo on black background next to text that says