Study: AI chatbots got about 1 in 7 voting questions wrong or out of date

Seven popular AI chatbots gave an outdated or wrong fact in 13.9% of answers to questions about how, when and where to vote, a study released before Election Day found.

The study was done by Dewey Square Group, a political consulting firm, for PSG Consulting and Innovating for the Public Good, a nonprofit.  It tested ChatGPT, Claude, Microsoft Copilot, Google's Gemini, AI Overviews and AI Mode, and Perplexity across 63 closely contested races.  About 5.9% of voting answers could have misled a voter, it said.

Which ones erred most

Anthropic's Claude and Copilot made three-quarters of the clear errors.  On voting questions, Claude was wrong about 1 time in 3, Copilot 1 in 4 and ChatGPT 1 in 10, the report said.

Answers in Spanish fared worse.  Thirty-two percent were refusals, errors or non-answers, against 14% in English.  About 1 in 10 links the bots cited came from low-quality sites such as betting-odds pages.

What they got wrong

The report grouped the mistakes into types that came up again and again, often for weeks:

• Voting rules:  a deadline that was still open was described as closed; Michigan's Election Day registration was placed at the polling place instead of the local clerk's office; and New Hampshire's repealed no-ID affidavit was offered as a fallback.
• Outdated links:  44 of ChatGPT's 57 voting errors were the same old web address for Texas's voter lookup page, since replaced by the state's My Voter Portal.
• Candidates not running:  withdrawn or defeated candidates were listed on the ballot or in polls.
• The wrong person:  a candidate was given the biography of someone else with the same name.
• Made-up statistics:  bill counts, attendance rates and letter grades from vote-tracker sites were stated as fact.
• Wrong numbers:  a district's past presidential margin was given for the wrong party or by the wrong amount, and a senator's confirmation votes were stated backwards.

Claude gave an answer that could mislead a voter 16.4% of the time on voting questions, against 0.2% for Google's AI Mode, the report said.

How that compares with other questions

The same study also graded every answer about 56 of the races, on topics such as projected winners, policy positions and candidate backgrounds, from Oct. 12 to 17.  Of those 22,203 answers, 5.1% had at least one clear factual error, and 10.9% had a clear error or a misleading claim such as a stale fact.  By comparison, 13.9% of 3,054 voting-logistics answers from Oct. 5 to 18 had an outdated or wrong fact.  The two figures come from separate grading passes with different definitions, so they are not an exact match.

Copilot and Claude led on the other questions too:  of the 1,127 answers with at least one hard error, Copilot gave 496 and Claude 330, and the other five platforms 301 combined, the report said.

Chatbots have also stumbled on general news.  A 2025 study by the European Broadcasting Union and the BBC, with 22 public media organizations in 18 countries, reviewed more than 3,000 answers from ChatGPT, Copilot, Gemini and Perplexity.  It found 45% had at least one significant issue, 31% had sourcing problems and 20% had major accuracy problems, such as made-up details or outdated information.  That study used different questions and grading, so its rates can't be compared directly with the election study's.

Polls in California close at 8 p.m. today.  Santa Clara County voters can check locations at the county Registrar of Voters, vote.santaclaracounty.gov.

Source:  Dewey Square Group, "AI and the Midterms, Part 2," October 2026, for Innovating for the Public Good and PSG Consulting; European Broadcasting Union and BBC, news integrity study, Oct. 22, 2025