NPR reported on October 30, 2026, that its test of ChatGPT found at least one site from a network of candidate-promoting websites in 19 of 37 responses. The test ran October 22–23. In 17 of those 19 responses, a cited site presented itself as local news. The findings came just days before the U.S. midterm elections scheduled for November 3.
What NPR found in ChatGPT’s citations
NPR tested 37 prompts in ChatGPT, Gemini and Google Search. It reported that ChatGPT cited sites from a network of 37 websites promoting Democratic candidates in 19 responses; 17 of those responses cited a site that presented itself as local news. NPR also reported that Gemini cited none of the network’s sites in its responses and declined to answer seven questions. Google Search did not surface the network on its first results page, and Google’s AI Overview did not mention it in the test.
Those results describe NPR’s test, not every response from these products. OpenAI spokesperson Nate Evans told NPR that ChatGPT’s search tools draw on a range of sources and that the company acts to improve responses when issues are brought to its attention.
What Searchable’s study measured
A separate test examined false answers to local political questions. Searchable says it graded 7,449 answers from ChatGPT, Gemini, Google AI Mode and Perplexity against official government data, describing the overall result as roughly one false answer in 12 (Searchable’s report). Tom’s Guide reported the detailed results below: 610 answers were classified as false, an overall rate of 8.2%.
The test covered 106 underlying questions across five metropolitan areas and their respective states. Tom’s Guide reported that the testing took place in September and October 2026. The rates apply to those questions and locations, rather than serving as general accuracy scores for every election question or product version.
| AI system | False-answer rate in the reported test |
| Perplexity | 22.6% |
| Gemini | 6.2% |
| ChatGPT | 4.4% |
| Google AI Mode | 2.9% |
Tom’s Guide also reported that about 62% of the 43,079 webpages accessed during the test were federal, state or local government pages. Access to official pages did not prevent false answers. The reported false-answer rate was more than four times as high for questions containing false assumptions as for questions without them.
How to check an election-related chatbot answer
A citation is a place to start, not proof that an answer accurately represents its source. Open the cited page and check whether it supports the specific claim, including its date and location. A page on the same general topic may not support the answer’s exact details.
Be especially careful with questions that build in an assumption. Searchable’s reported test found a much higher false-answer rate for prompts containing false assumptions, so phrase the question neutrally rather than treating a disputed claim as established. For polling places, registration and ballot details, check with the relevant state or local election office.
