ChatGPT, Gemini and Perplexity are getting election facts wrong — and a new study shows why
Latest from Tom's Guide ·
null
With midterm elections just days away, you might be tempted to ask ChatGPT, Gemini, or Perplexity about a candidate’s record before heading to the polls. But a new study suggests you should think twice before trusting the answer, even when the chatbot provides sources to back it up.
A report from AI search analytics company Searchable found that roughly one in 12 answers about local political issues contradicted official government data. Of the 7,449 answers researchers evaluated, 610 were false, an 8.2% error rate.
The research examined four popular AI tools: ChatGPT, Gemini, Google AI Mode , and Perplexity. The questions covered issues voters regularly consider when evaluating politicians, including crime, housing costs, employment , education and healthcare.
But the most surprising finding wasn’t necessarily how often the chatbots got things wrong. It was that they sometimes produced incorrect answers despite consulting reliable government sources.
The researchers found significant differences in how the four AI tools performed.
Perplexity produced false answers in 22.6% of the tested responses, compared with 6.2% for Gemini and 4.4% for ChatGPT. Google AI Mode had the lowest error rate at 2.9%.
In other words, more than one in five Perplexity answers evaluated in the study contradicted official government data.
Errors also varied by subject. Questions about housing and the cost of living produced the most false answers overall, at 11.9%, followed by energy and environmental issues at 10.9% and schools at 10.1%.
Crime and public safety questions had the lowest aggregate error rate, at 4.7%.
These figures don’t represent each chatbot’s correctness across all political questions. The study focused on 106 underlying questions across five metropolitan areas and their respective states, including questions deliberately designed around misleading or previously debunked claims.
One of the more concerning findings is that incorrect answers weren’t necessarily the result of chatbots consulting unreliable websites.
Of the 43,079 webpages the AI systems accessed during the study, approximately 62% came from federal, state, or local government websites.
ChatGPT was particularly thorough when gathering information. Researchers found that it consulted an average of 31 webpages from 10 websites before answering a single question.
One example shown in the report involves Georgia’s electricity supply, ChatGPT reportedly ran 19 searches and accessed 154 pages across 52 websites before responding.
That highlights a problem with relying on AI to interpret information for you. A chatbot can search extensively, access authoritative sources, and still make an error when pulling the information together.
As Searchable CEO and co-founder Chris Donnelly explained, “Often the right answer was sitting in the sources the chatbot linked to; it just didn’t make it through to the voter.”
The concern becomes especially relevant when an answer contains a statistic that could influence how someone evaluates an elected official’s record.
The study also examined what happens when questions contain a misleading assumption.
For example, asking a chatbot why crime increased in a particular area already assumes crime went up, even if the official statistics show otherwise.
According to Searchable’s findings shared with Tom’s Guide , questions containing false assumptions were more than four times as likely to produce a false answer.
AI doesn’t always challenge a question's premise. Sometimes it accepts that premise and generates an explanation around it.
This could be particularly problematic for people researching political claims they’ve encountered on social media, in campaign advertisements, or during debates.
And Americans are already using AI for that purpose. An October survey published by George Washington University’s Graduate School of Political Management found that one in five likely midterm voters had asked AI about elections or politics and a quarter reported using AI every day.
Interestingly, the same survey found that just 1% considered AI their most trusted source for information about elections and candidates.
Obviously, none of this means you need to stop using ChatGPT, Gemini, or other AI tools to research political issues. They can still help explain complicated policies, summarize information, or find relevant sources.
But take a few precautions before accepting an answer as fact:
AI chatbots have made finding information easier, but this study shows why finding information and interpreting it correctly are two different things.
Searchable tested its systems in September and October 2026, using anonymized logged-in accounts and repeated questions across fresh sessions. Researchers checked the answers against official government data, and two graders had to agree before they classified an answer as false.
The study isn’t a comprehensive measure of chatbot reliability, but it does illustrate a weakness worth understanding before the November 3 midterm elections.
As Donnelly put it, “Treat a chatbot’s answer as a starting point, not a verdict.”
With so much information available in seconds, spending an extra minute checking a source could be the most useful part of your research