The numbers Anthropic is working from
Anthropic has set out what it is doing to Claude ahead of the 2026 US midterms, and the post opens with measurement rather than policy. In a random sample of about 250,000 US Claude.ai conversations between 21 and 27 September, the company says roughly 5.75% touched on politics, public policy or social issues; about 0.51% were specifically about elections or voting; about 0.01% asked whom or what to vote for; and about 0.002% violated its election integrity policies.
That last figure is small, and Anthropic expects all of them to rise closer to election day. The post is the company’s own analysis of its own logs, and none of it has been independently audited.
Banned accounts, named narratives
The sharpest part of the post is not the product work. Anthropic says foreign operators used Claude to produce US midterm coverage favouring Russian, Iranian and North Korean narratives, and that the accounts were banned. Separately, it says threat actors used Claude to generate fabricated news for ad revenue; those accounts were banned too and detection was improved.
Anthropic also describes a gap its own testers found. Working with the Institute for Strategic Dialogue and the McCain Institute, it stress-tested scenarios across the United States, Israel and Brazil. Claude reliably refused single harmful requests — but could miss a campaign broken into tasks that each look innocuous on their own. New classifiers were added in response, aimed at multi-step misuse, fake accounts and fake news sites built to seed political discussion, financially motivated election schemes, and foreign state actor activity.
What changes for users
Later this month, Claude will pull official voting information from Democracy Works into its US answers, so that a question about where to vote or what identification to bring gets a state-specific reply. The TurboVote banner introduced in April stays. On election night and afterwards, Claude will show up-to-date official results.
On political even-handedness, Anthropic reports that Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 all score 99% on its own evenhandedness evaluation, and that newer models engage more readily and decline fewer requests. Three outside groups — The Future of Free Speech, the Foundation for American Innovation and the Collective Intelligence Project, which surveyed more than 2,000 people in the US, UK and India — found Claude rarely refuses but sometimes adds caveats nobody asked for, and that Claude Sonnet 4.5 hedged less evenly on right-leaning arguments. Anthropic says it will work on yes-or-no answers, on when caveats get added, and on widening bias tests to cover candidates and voting rather than only policy.
The rulebook underneath
All of this sits on a usage policy update published the previous week, which takes effect on 12 November. It adds a deceptive activity section covering fake accounts, fabricated news sites and influence operations, political or commercial; renames the elections section “Do Not Undermine Democratic Processes”; and drops the blanket ban on personalised vote and campaign targeting. A separate guide for campaigns lists what is allowed — press statements, debate preparation, voter outreach — and what is not, including fake personas, inflated engagement and content presented as independent news when it is not.
The checkable thing from here is the multi-step gap. Anthropic has published that single harmful requests get refused and distributed campaigns sometimes do not. Whether the new classifiers close that is something only the next enforcement report will show.
