AI Chatbots Supply False Answers on Local Issues and Voting Rules Ahead of US Midterms

AI chatbots are providing incorrect information on local political issues and voting procedures in the final days before the US midterm elections on 3 November. One study that graded 7,449 answers from ChatGPT, Gemini, Google AI Mode and Perplexity against official data found that 1 in 12 answers, or 8.2 percent, stated something false. Questions on the cost of living and housing produced the highest error rate at 11.9 percent. Perplexity answered falsely more than 1 in 5 times, or 22.6 percent. Google AI Mode was wrong least often, at 2.9 percent.

A separate analysis by Forum AI examined answers to election-related queries and found that almost 40 percent contained at least one factual error. When researchers asked Anthropic's Claude why voters should elect a specific North Carolina representative, the model listed a credential belonging to a different person. Google's Gemini gave an incorrect deadline for requesting a mail-in ballot in Pennsylvania, stating 3 November when the actual cutoff was a week earlier. A voter following that answer would have applied too late.

The Washington Post tested more than 3,000 AI-generated voter guides. Chatbots frequently invented races that were not on the ballot, misidentified which congressional district an address belonged to, and listed local officials as candidates for federal office. They failed to name the correct House district 48 percent of the time. When they did name a district, they were wrong 22 percent of the time. Claude and Alexa+ produced the highest error rates in that test, each incorrect for more than 1 in 5 addresses.

Researchers noted that chatbots often cited low-quality sources, including Reddit threads, and sometimes attributed statements to candidates or officials that those individuals had not made. OpenAI and Anthropic have announced partnerships intended to supply trusted election information, yet the tested models still produced significant errors on logistics and candidate details. Millions of voters are now using these tools for political information in the first federal election cycle in which such reliance is widespread.

The studies focused on questions voters use to judge records on crime, jobs, schools, energy and living costs, as well as basic voting rules. Error rates varied by model and by topic. Housing and cost-of-living questions generated more false answers than some other categories. Voting-procedure questions produced incorrect deadlines and procedures more often than questions about candidates' biographies in some tests.

No federal agency has issued a comprehensive public tally of chatbot-related voter confusion. State election offices have published their own guides and warned voters to consult official sources. The chatbot errors documented in the studies include both factual mistakes about candidates and procedural mistakes that could prevent a valid ballot from being cast. The scale of use is not measured in the reports, but the researchers treated the tools as already active in the information environment.

The midterms on 3 November will test whether voters who consult chatbots cross-check the answers against state and local election sites. The published error rates indicate that a non-trivial share of responses on local issues and voting logistics do not match official records. Models differ sharply: one study placed Perplexity above 20 percent false answers while another model stayed under 3 percent. Users receive no consistent signal of which system is more reliable on a given query.

The findings arrive as campaigns and election administrators continue to direct voters to official channels. The documented mistakes range from swapped candidate credentials to invented contests and late ballot deadlines. Whether those errors change any votes, or merely add noise, will not be known until after the ballots are counted. The studies supply concrete counts of incorrect statements against verifiable public data in the final week before polling.