Anthropic's midterms plan: only 0.01% of sampled Claude chats asked whom to vote for

How will Claude handle the 2026 US midterms? Anthropic's answer, published October 14, 2026, is a mix of usage data, outside testing and a list of the abuse it has already seen. The headline number is small: in a one-week sample of 250,000 US consumer conversations, about 0.01% asked Claude whom or what to vote for.

That figure and most others in the post are Anthropic's own. They deserve careful reading, not dismissal, because the company also published the places where outside reviewers found problems.

What people actually ask

Anthropic analyzed 250,000 random US Claude.ai consumer conversations from September 21 to 27, 2026. About 5.75% touched on politics, public policy or social issues. About 0.51% were about elections or voting. About 0.01% asked for a voting recommendation. About 0.002% violated its election integrity policies.

Most political queries, the company says, were informational: researching candidates or weighing policy arguments. It expects these shares to grow closer to Election Day. This is one week, one product and consumer traffic only. API users, including campaigns building their own tools, are not in the sample.

Voting information and results

A TurboVote banner from Democracy Works, a nonpartisan resource, continues to appear for registration, polling place and election date questions. Later in October, Claude will pull official, state-specific voting information from Democracy Works for US users, including where to vote and what ID to bring. On election night and in the days after, Claude will show the latest official results.

Routing to an official source is a sound design choice. It avoids the model improvising a polling address, which is exactly the kind of error that suppresses turnout.

The 99% evenhandedness claim

Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 each scored 99% on Anthropic's political evenhandedness evaluation, which the company calls its most evenhanded results to date. Three outside groups reviewed the models: The Future of Free Speech at Vanderbilt University, the Foundation for American Innovation and the Collective Intelligence Project.

The reviewers' findings are more useful than the score. Claude rarely refused requests but sometimes added unsolicited caveats or counterarguments. The Foundation for American Innovation compared matched left- and right-leaning prompts and found Claude Sonnet 4.5 did not always hedge evenly and was more likely to soften right-leaning arguments. Anthropic says it will improve balance on short-answer questions, make caveats more consistent and broaden bias tests to cover candidates and voting.

Note that the model flagged in that finding is an older one. Anthropic does not say the current models repeat the pattern, and a 99% self-reported score cannot settle it. A third party running the same prompts on today's models would.

Red teaming and what slipped through

The Institute for Strategic Dialogue and the McCain Institute tested safeguards in the US, Israel and Brazil. Claude reliably refused single harmful requests but could miss a malicious campaign split into harmless-looking tasks, such as audience research, building a news site and drafting content.

Anthropic added classifiers for multi-step misuse, fake accounts and news sites, financially motivated election schemes and foreign state actors, plus localized safeguards for the three countries. Confirmed violations lead to account bans.

Abuse already seen

Foreign-based operators used Claude to produce US midterm coverage favoring Russian, Iranian and North Korean narratives. Anthropic banned the accounts. Other actors used it to fabricate news for ad revenue, and those accounts were banned too.

The Usage Policy, updated on October 8 and effective November 12, narrows the election section to disallowing "the use of Claude to deceive voters or disrupt elections," and removes a blanket ban on personalized vote and campaign targeting so that legitimate civic work, such as nonprofits writing voter information in other languages, is allowed. Deceptive targeting stays prohibited under other sections. We covered the full change in our usage policy report.

Anthropic also published a campaign guide. Permitted uses include press drafts, debate prep and turnout analysis. Prohibited ones include fake personas, inflated engagement and content posing as independent news.

What to watch

Anthropic gave no figures for how many accounts it banned or how many flagged operations it found, so the scale of the abuse is unknown. It also published no data on the multi-step classifiers' false-positive rate, which matters because legitimate campaign staff do multi-step work all the time.

How to read the numbers

Three cautions apply. First, the 0.002% policy-violation rate is a share of sampled consumer chats, and it only counts what Anthropic's own detection recognized. Second, a one-week window in late September predates the heaviest campaign period, which Anthropic itself expects to grow. Third, the evenhandedness score comes from an evaluation Anthropic designed, even though outside reviewers then examined the models against their own prompts.

None of that makes the post unreliable. It means the figures are best read as Anthropic's disclosure, to be compared with whatever other labs publish and with an outside rerun.

Our take

The most believable sentence in the post is the admission that single-request refusals are not enough. Election abuse is a campaign problem, and campaigns are made of boring tasks. The classifiers for multi-step misuse are the right fix, and also the hardest to verify from outside.

Our position: read the 0.01% as reassurance about what ordinary people ask, not about what organized actors do. The numbers to check after November are the ban counts and an independent rerun of the evenhandedness test on the current models.