Anthropic set out on October 14, 2026 how Claude will handle the US midterm elections, combining new voter-information features, the results of outside bias reviews and tighter abuse detection. The post builds on an election-safeguards update the company issued earlier in the year and on the revised Usage Policy it released the previous week.
The company first measured how Americans use its consumer app for politics. Using its Anthropic Insights tool, it drew roughly 250,000 Claude.ai chats from US users at random for the week of September 21 to 27, 2026, and classified them with a method it describes as privacy-preserving. Politics, public policy and social issues came up in about 5.75% of them, and elections or voting specifically in about 0.51%. Requests asking Claude whom or what to vote for were around 0.01%, and a separate review of another 250,000 chats found policy-violating election requests in about 0.002%. Anthropic expects the political share to rise as Election Day approaches.
Two features are being added for US users of Claude.ai. Later in October, answers about where to vote or which ID to bring will draw on official data compiled by Democracy Works, a nonpartisan group that gathers information from state and local election offices, so the reply can be specific to a user's state. Claude will also display current official vote counts on election night and in the days after. The existing banner pointing to Democracy Works' TurboVote service stays in place.
On bias, Anthropic says the Claude 5.5 family (Opus, Sonnet and Haiku) scores 99% on its internal political even-handedness evaluation and refuse fewer political requests than earlier versions. It also commissioned three outside reviews. The Future of Free Speech, a think tank at Vanderbilt University, tested hundreds of contentious prompts; the Foundation for American Innovation compared answers to mirrored left- and right-leaning requests; and the Collective Intelligence Project built an evaluation from principles suggested by over 2,000 participants across the US, the UK and India. Anthropic reports that most responses were rated strong, but the reviewers also flagged unrequested caveats and counterarguments, and the Foundation for American Innovation found Claude Sonnet 4.5 more inclined to soften arguments for right-leaning positions. The company says it is adjusting short and yes-or-no answers and how it adds caveats.
The most concrete security finding came from red-teaming by the Institute for Strategic Dialogue and the McCain Institute in the US, Israel and Brazil. Testers found Claude consistently declined a single overtly harmful request, yet could miss a deceptive campaign when the work was broken into small tasks such as audience research, building a news site, drafting posts and planning distribution. Anthropic says it responded with classifiers aimed at multi-step and novel misuse, including fake social media personas, sham news outlets and election schemes run for profit, plus better detection of foreign state actors and stronger cyber protections for campaigns and election infrastructure.
Enforcement examples are included. Anthropic says a fake news site run from abroad relied on Claude to produce midterm stories that favoured Russian, Iranian and North Korean narratives, and that other actors generated fabricated stories to draw American readers and advertising income. The accounts in both cases were banned. The company has also published a guide for political campaigns describing permitted uses, such as drafting press releases or analysing turnout data, and banned ones, such as fake profiles or inflated engagement.
Most figures in the post, including the usage shares and the 99% even-handedness score, come from Anthropic's own measurements, and the outside reviewers' findings are presented in Anthropic's summary rather than in full reports. The company says it plans a fuller account of election-season usage in the coming months.