AI chatbots are getting practical election questions wrong often enough to matter, and the gap is markedly worse in Spanish. A study reported by the Latin Times tested six chatbots on 2,400 voter questions across ten US states and found that even answers in English were incomplete, outdated or potentially misleading 29% of the time.
Watch Desk analysis
What happened
The Institute for Strategic Dialogue tested GPT-5.5, Gemini 3.5 Flash, Muse Spark, Grok 4.3, DeepSeek V4 Pro and Sonnet 4.6, with web search enabled. The report says complete, accurate answers fell from 71% in English to 55% in Spanish. The problem was usually omission rather than outright invention: a chatbot might provide the basic rule while dropping the deadline, exception or backup option that makes it useful.
The reported spread between models was substantial. GPT-5.5 scored 89% for complete, accurate English answers in the tested set, while Muse Spark's Spanish accuracy was 38%. On the question “Can I vote by mail?”, Spanish answers were complete only 42% of the time. These are findings from June testing, not a live verdict on today's app versions.
Key findings
- English answers still fail
Nearly three in ten English responses were judged incomplete, outdated or potentially misleading. - Spanish answers lose vital detail
Overall complete, accurate performance fell to 55% in Spanish, with omissions often carrying the practical risk. - Model quality varied
GPT-5.5 led the tested group in English, while Muse Spark performed particularly poorly in Spanish. - Voting by mail was weakest
Spanish answers to this ordinary voter question were complete just 42% of the time. - Conspiracy refusals held up
Refutation rates for debunked voting conspiracies were broadly similar across English and Spanish. - The test has boundaries
Five systems were accessed through OpenRouter, and newer models may behave differently today.
Why it matters
A missing polling deadline or identification rule is not a charming conversational glitch. It can turn a technically plausible answer into unusable guidance, particularly when a voter asks in Spanish and receives less of the detail needed to act.
The practical rule is simple: use a chatbot to find the shape of a question, then verify deadlines, ID requirements, polling locations and mail-ballot procedures with the relevant state or county election office. The study does not prove that every chatbot will fail every voter, but it does show why fluent prose is a poor substitute for verified election information.
Our read
This is a useful warning about completeness, not merely factuality. The most dangerous answer may be the one that sounds right while quietly leaving out the exception. Developers should test multilingual answers for the details people need to act on, and voters should treat chatbot guidance as a starting point rather than an authority.
What to watch
- Whether newer model versions close the English-Spanish completeness gap.
- Whether consumer apps perform differently from the OpenRouter tests.
- Whether election platforms route users to official voting information for deadlines and local rules.
- Whether independent testing expands beyond the ten states and includes major Spanish-speaking population centres.
Discussion spark: Should election chatbots be required to hand off deadline, ID and mail-ballot questions to official state or county sources?
Sources and evidence
- AI Chatbots Get Nearly 1 in 3 Voting Questions Wrong – And the Errors Climb Higher in Spanish – Latin Times (12 September 2026, 19:58 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.