On October 14 Anthropic published what it is taking into the US midterms. Two things are new in the product. The rest is measurement, and the measurement is the more interesting half.
What 250,000 conversations show
Using Anthropic Insights, one week in September was classified. Conversations that touch politics, public policy or social issues make up about 5.75 percent of all US Claude.ai conversations. Elections or voting specifically account for 0.51 percent. Asking Claude whom or what to vote for shows up in roughly 0.01 percent. Conversations that violate the election integrity policies sit at 0.002 percent.
Democracy Works supplies the official answers
Until now, questions about registration, polling places or election dates produced a banner linking to TurboVote. Later this month Claude starts pulling Democracy Works data straight into its replies: ask where to vote or what ID to bring, and you get a state-specific answer without leaving the conversation. On election night and the days after, Claude will also show the latest official vote counts.
Three outside groups tested the balance
Opus 5.5, Sonnet 5.5 and Haiku 5.5 score 99 percent on Anthropic’s own political evenhandedness evaluation. Three independent groups ran their own: The Future of Free Speech at Vanderbilt University, the Foundation for American Innovation with matched prompts from left and right, and the Collective Intelligence Project with more than 2,000 people surveyed in the US, UK and India. Two findings sting. Claude rarely refuses, but it adds unsolicited counterarguments — dropping an opposing view into an advocacy email, for instance. And Sonnet 4.5 softened right-leaning arguments more often than left-leaning ones.
The campaign made of harmless pieces
The sharpest result came from red-teaming with the Institute for Strategic Dialogue and the McCain Institute, run in the US, Israel and Brazil. Claude reliably refused a single harmful request. Then the analysts broke the same campaign apart: research the audience, build a news site, draft the content, plan the distribution. Every step looked innocuous on its own, and that is exactly how a campaign can get through. New classifiers now target that multi-step pattern, alongside better detection of foreign state actors.
The backdrop fits. According to an analysis shared exclusively with Axios, AI safety groups and individual donors have spent nearly $50 million this election cycle on candidates who favour regulation. One name in that mix is Coefficient Giving, a philanthropy with close ties to Anthropic executives.
Single requests are not the problem
What stands out in the numbers is how small everyday misuse is: 0.002 percent. The real gap sits somewhere else, in the split-up campaign, and each individual conversation in it looks legitimate. A classifier built to catch that has to reason across conversations, which is exactly where product safety and privacy start pulling in opposite directions. That Anthropic published the finding instead of leaving it out is the most useful part of the paper.