Do AI chatbots give voters accurate election information?

States United Democracy Center ran two rounds of empirical testing of ChatGPT and Google AI on common voter questions, in late 2025 and early 2026, ahead of the 2026 midterms. In the "preliminary" round, error rates were 6.9% (Google AI) and 8.2% (ChatGPT); by the "primary" round, "the rate of verifiable factual errors fell to 0% across both platforms." The study's bottom line, though, is that this accuracy gain did not make the platforms adequate substitutes for official sources: "voters seeking election information on leading AI platforms are inconsistently directed to the authoritative information on their state's official election website." ChatGPT mentioned a state election site in 39.4% of responses and Google AI in 55.6% — "less than 50% of the time" overall, and "for some questions, they almost never did."

The report is explicit that zero measured errors is a narrower claim than it sounds: "It is worth being precise about what a 0% error rate does and does not establish... It does not mean, however, that the responses were complete." ChatGPT returned incomplete candidate lists for 88.9% of gubernatorial queries, which the study attributes to a structural mismatch — "AI systems trained on historical data and reliant on periodic retrieval were structurally ill-suited to track real-time candidate filings" like an actively changing field such as Arizona's. Compounding this, candidate questions relied heavily on Wikipedia (12.3% of all 3,481 cited links, and the dominant source for candidate questions specifically), which the study flags as "openly editable and not authoritative for time-sensitive or jurisdiction-specific election information." Google AI also changed its output format mid-study, on or around Feb. 2, 2026, replacing prose summaries with links-only answers; asked why, Google attributed the change both to reducing hallucinations and to "making AI search more monetizable" — after which Google AI directed voters to a state site in none of its responses.

This sits alongside Did ChatGPT cause Stack Overflow posting to decline? as a parallel case of a chatbot interposing itself between a user and the authoritative source rather than routing to it — there, a drop in traffic to the platform that held the answers; here, a measured shortfall in links back to the office that administers the information. It also echoes How much GPT-written scholarship reaches Google Scholar undetected? in flagging degraded provenance in a nominally authoritative channel, though the two findings are not the same mechanism: Haider's concern is undisclosed AI content entering a scholarly index, while States United's is a chatbot citing an editable encyclopedia and unvetted platforms (Reddit, YouTube) for facts that only an election office can certify.

The study covers two AI platforms, a defined question set, and three states over several weeks in round two, so it does not establish how other assistants, other states, or other election cycles would score, nor does it isolate how much of the completeness gap is retrieval architecture versus deliberate design choice (the "monetizable" framing for Google's format change is Google's own stated explanation, not something the study verifies independently). Within those bounds, the finding supports treating chatbot accuracy and chatbot completeness as separate, non-substitutable measures of voter utility — a system can reach zero measured factual errors while still failing the single measure the study treats as most consequential, directing the user onward to the source that is actually accountable for the answer.