For two election cycles, Google took the cautious route. In 2022 and 2024 it blocked Gemini from answering many election-related questions and pushed users toward Google Search instead.

This week it changed tack.

TIME reports that Google will add election information about the upcoming US midterms to the Gemini app, and to AI-generated responses across its search products. Specifically: polling locations and voter registration instructions drawn from state and local governments and the nonprofit voting-information group Democracy Works, plus race results from The Associated Press.

"People come to Google to stay informed during election season—like when finding their polling site, watching candidate debates, or tracking results on election night," the company said in its blog post. "We take the responsibility to deliver high quality, trustworthy information seriously."

It's worth noticing that this design is more conservative than the headline suggests. Google is not turning a language model loose on election law. It is routing questions to named authoritative sources. That is the right architecture.

Google also isn't alone. OpenAI and Anthropic have both said their chatbots will direct users to Democracy Works. OpenAI additionally plans to provide live vote counts from the AP and to monitor its systems for signs of political bias. Meta spokesman Corey Chambliss said MetaAI will give users local information or direct them to government sources when they ask about voting.

The whole industry is opening this gate at once.

A week earlier, the Institute for Strategic Dialogue published research finding that chatbots from major companies — including OpenAI, Google and Anthropic — routinely gave incorrect answers to questions about voting. The problem was worst in states that had significantly changed their election procedures.

The numbers: 29% of responses to English prompts about voting were either incomplete, inaccurate, or outdated. Responses in Spanish were 16% less likely to be accurate.

"We know that voters are going to turn to chatbots more and more," said Valeria de la Fuente, a digital research analyst at ISD who co-authored the report. "So the quality of the responses that we found is concerning."

This is the part that deserves the most attention, because the failure is not the one people expect.

It wasn't invention. It wasn't bias. It was staleness.

"The AI systems appeared to be pulling outdated information from state government websites about the 2024 election," TIME reports. The models went to the right kind of source — an official state election page — and read the wrong year.

The questions that failed were not exotic. ISD listed some of the lowest-performing ones:

These are exactly the questions a real person asks four days before an election, when the answer actually determines whether their vote counts.

And a stale answer is the most dangerous kind, because it has every marker of a good one. It is specific. It is fluent. It cites a government website. It is simply describing rules that no longer apply.

The Spanish-language results were worse in a way the percentage alone doesn't convey.

"Overall, what we saw was a lot of ambiguous, confusing language" in the Spanish responses, de la Fuente said. "In some cases, poor or neutral translations from English. Some of them could still be understood. Some of them were simply wrong."

TIME notes the responses in Spanish were also often less detailed, or more likely to contain no citations at all — reflecting, de la Fuente said, the language disparities in the overall information ecosystem.

That last point matters. The chatbot gap is downstream of a real-world gap. There is less good Spanish-language election information on the internet, so there is less for a model to retrieve. This will not fix itself as models improve, because the shortage isn't in the model.

One more finding, and it's the one that has stayed with me.

The models were good at answering adversarial prompts raising specific election fraud claims that media organizations had already fact-checked. Those have clean, well-documented debunks sitting in the training data.

They were weaker at harder-to-refute election-related concerns — ballot harvesting, voting machine security, noncitizen voting.

"It's in those 'gray areas' where the models are more likely to give 'problematic' answers," de la Fuente said.

Read that as a competence curve running the wrong direction. These systems are strongest where the answer was already easy to find, and weakest exactly where a person is most likely to be genuinely confused and most in need of careful, sourced help.

(To be clear about what's being said: ISD's finding is about how well the models handled these topics, not a verdict on the underlying disputes.)

A Pew Research Center study earlier this year found about half of adults under 50 rely on chatbots to search for information, while a smaller share use them to read news.

That is the number that turns a technical accuracy finding into a civic one. A 29% error rate on a niche tool is a research curiosity. The same error rate on a primary information gateway for half of everyone under 50 is something else.

Not a return to blocking. Google's 2022–2024 posture — refuse and redirect — was not obviously safer. People who don't get an answer go looking elsewhere, and "elsewhere" during an election is not reliably better. Refusal isn't neutrality.

Publish the error rate. If these products are becoming a civic information layer, accuracy on voting procedure is a public-interest metric. Right now the only people measuring it are outside researchers doing it unpaid and uninvited. The companies are better positioned to test this continuously than ISD is, and none of them publish results.

Fix freshness before capability. The failure here wasn't a lack of intelligence. A more capable model that reads the same outdated page gives the same wrong answer with more confidence.

For voters, one concrete habit: use a chatbot to orient, then confirm anything procedural — deadlines, ID requirements, absentee documents, polling location — with your state or county election office directly. The chatbot's own cited source is usually the place to go; just make sure you're reading the current page and not the model's memory of an old one.

For parents and teachers: this is a good, low-heat example for teaching source-checking, because nothing about it is partisan. The lesson is mechanical — a confident, well-written, government-citing answer was still wrong, because it was a year out of date. That generalises far beyond elections, and a student can verify it themselves in about two minutes against a state election site.

Disclosure: Anthropic, the company that makes the AI model used to help assemble this newsletter, is one of the three companies whose chatbots the ISD study found giving incorrect answers. It seemed better to say so than to leave it out.

Source: TIME, reporting by Naomi Nix: https://time.com/article/2026/09/10/election-voting-gemini-chatgpt-claude/

Want practical AI guidance for parents and educators every week? Subscribe: https://www.aibyage.com/?modal=signup&utm_source=beehiiv&utm_medium=newsletter