# AI Misinformation Report: US Midterm Elections 1 in 12 answers AI chatbots gave to questions about local political issues were false. By Connor Lahey (Content Lead) and JP Garbaccio (Head of AEO/SEO). Updated October 7, 2026. Ahead of the midterm elections on November 3, we asked ChatGPT, Gemini, Google AI Mode and Perplexity 106 questions about the local issues voters use to judge a politician's record, from crime and jobs to schools, energy and the cost of living. We graded 7,449 of their answers against official government data, and sorted the 43,079 web pages the chatbots read while answering. ## Key findings - Of 7,449 answers to questions about local political issues, 1 in 12 (8.2%) stated something official data shows is false. - Perplexity answered falsely more than 1 in 5 times (22.6%). Google AI Mode was wrong least often, in 2.9% of answers. - Questions about the cost of living and housing were answered falsely most often (11.9%), followed by energy and the environment (10.9%) and schools (10.1%). Why it matters: when AI chatbots change people's political views, they do it with facts. Two studies published in December 2025 found that chatbots set up to argue for a candidate shifted voters' preferences by more than traditional video ads usually do, mainly by citing facts and evidence (Nature, https://www.nature.com/articles/s41586-025-09771-9), and that the methods that made chatbots more persuasive also made their claims less accurate (Science, https://www.science.org/doi/10.1126/science.aea3884). ## How often do AI chatbots give false answers about local political issues? AI chatbots contradicted official government data in 610 of 7,449 answers, or 1 in 12 (8.2%). A false answer here means the chatbot told the user something the official figures contradict: that a number rose when it fell, that a state sat above the national average when it sat below, or that something happened when it didn't. These are the facts voters use to judge an incumbent, such as whether crime went up on their watch, whether jobs came back, or whether homelessness levels improved. When a chatbot gets one of them wrong, the user can end up crediting or blaming a politician for something that never happened. ## Which AI chatbot gives the most false answers about local issues? | AI chatbot | False answers | Answers | |---|---|---| | Perplexity | 22.6% | 334 of 1,479 | | Gemini | 6.2% | 133 of 2,153 | | ChatGPT | 4.4% | 96 of 2,168 | | Google AI Mode | 2.9% | 47 of 1,649 | Perplexity gave a false answer more than 1 in 5 times, over three times the rate of the next chatbot, Gemini, and supplied more than half (334) of the 610 false answers. The type of fact each chatbot got wrong most often varied. Perplexity's highest false-answer rates were on the cost of living and housing (30.8%) and on energy and the environment (29.8%). Gemini was most often wrong about the cost of living and housing (13.0%) and schools (8.7%). ChatGPT's highest rate was on jobs and the economy (6.5%), and Google AI Mode's on energy and the environment (5.5%). Even the most accurate chatbot, Google AI Mode, gave a false answer 2.9% of the time. That is enough to mislead 1 in 35 people who ask it about a local political issue. ## Which local issues do AI chatbots get wrong most often? | Issue | False answers | Answers | Questions | |---|---|---|---| | Cost of living and housing | 11.9% | 84 of 703 | 8 | | Energy and environment | 10.9% | 109 of 1,002 | 14 | | Schools and education | 10.1% | 92 of 909 | 14 | | Jobs and the economy | 8.8% | 112 of 1,269 | 18 | | Population and immigration | 7.2% | 78 of 1,090 | 14 | | Health and healthcare | 6.6% | 80 of 1,205 | 17 | | Crime and public safety | 4.7% | 51 of 1,084 | 17 | The three issues at the top rest on figures that change every year, such as median income, rents, electricity prices and test scores. AI chatbots struggle to give accurate information on facts that change often, because an answer built from an older set of figures can be out of date by the time someone asks. ## How many sources do AI chatbots read before answering about local American issues? Average web pages read to answer one question about a local issue: | AI chatbot | Pages read per answer | Websites per answer | |---|---|---| | ChatGPT | 31 | 10 | | Perplexity | 9 | 7 | | Google AI Mode | 6 | 5 | | Gemini | 2 | 2 | ChatGPT ran 3.3 searches per answer on average. The most for a single answer was 19 searches and 154 pages from 52 websites, for a question about how much of Georgia's electricity comes from nuclear power. ## What types of website do AI chatbots read to answer questions about local issues? Share of the 43,079 web pages read, by type of website: | Type of website | All four | ChatGPT | Perplexity | Google AI Mode | Gemini | |---|---|---|---|---|---| | Federal government | 43.3% | 56.1% | 22.2% | 18.8% | 18.7% | | State and local government | 18.8% | 20.4% | 17.1% | 15.0% | 13.5% | | News outlets | 10.8% | 4.7% | 17.0% | 30.9% | 17.5% | | Data aggregators | 8.9% | 6.3% | 16.2% | 9.3% | 13.4% | | Think tanks and advocacy | 7.3% | 4.8% | 10.5% | 13.0% | 13.9% | | Company websites | 4.9% | 2.3% | 9.3% | 7.7% | 15.1% | | Universities and schools | 2.2% | 1.4% | 4.7% | 2.0% | 3.8% | | Other | 3.7% | 4.0% | 3.0% | 3.3% | 4.0% | Government websites made up 62% of the pages: 43.3% federal and 18.8% state and local. ChatGPT leaned hardest on government (76.5% of its pages). Google AI Mode read news outlets more than any other type (30.9%). Gemini read the least from government, under a third of its pages (32.2%), and the most from company websites, at 15.1%. ## Methodology - Searchable tested five metro areas and their states for this first edition: Atlanta (Georgia), Charlotte (North Carolina), Houston (Texas), Las Vegas (Nevada) and Portland (Maine). - We wrote 106 questions. 59 ask about a local fact with one answer in official data. 47 are built from fact-checked false claims that politicians and political groups on both the left and the right have made about those five states, each asked as a multiple-choice question with the false claim as one option. - Every correct answer was confirmed by two researchers working separately, against official government data only. - Questions were asked in several forms: open, misleading, multiple choice, multiple choice with a request to list sources, and a state's questions all in one prompt. Each was asked five times in fresh sessions, through anonymized logged-in accounts rather than the chatbots' APIs. - 7,449 answers were graded. An answer counts as false only when two graders agreed it stated a wrong answer. You are free to report on these findings with credit to Searchable and a link to this page.