Anthropic said on Oct. 14 that, later this month, its Claude chatbot will start answering U.S. users' voting questions with official information from Democracy Works, a nonprofit. Someone who asks where to vote or what ID to bring to the polls on Claude.ai, Anthropic's chatbot app, will get an answer specific to their state without leaving the chat. The answers draw on public information Democracy Works collects from state and local election sources. Anthropic says Claude will also show the latest official vote counts on election night in November and in the days that follow.
Politics makes up a small share of how people use Claude, Anthropic's own count shows. In a random sample of 250,000 U.S. conversations on Claude.ai from Sept. 21 to 27, about 5.75% touched on politics, public policy or social issues, and about 0.51% were about elections or voting. Requests asking Claude whom or what to vote for were rarer still, about 1 in 10,000 conversations.
Anthropic also tested how Claude holds up against people trying to use it to deceive voters. Anthropic asked two outside research groups, the Institute for Strategic Dialogue and the McCain Institute, to test its protections in the U.S., Israel and Brazil. The testers first asked for harmful help directly, then broke whole campaigns into small jobs that each look innocent, such as researching an audience, building a news site, writing the content and planning how to spread it. The testers found that Claude reliably refused a single harmful request, Anthropic says, but could miss a deceptive campaign when it was spread across many requests that each looked harmless on their own.
Anthropic says it has since added new automated filters meant to spot that kind of step-by-step misuse, including attempts to build fake social media accounts and fake news sites about politics. It says foreign-based operators of a fake news site used Claude to write midterm coverage favoring Russian, Iranian and North Korean narratives, and that others used Claude to make fabricated news to draw American readers and ad money. Anthropic says it banned the accounts involved.
On bias, Anthropic says its newest models, Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, score 99% on a political evenhandedness test the company designed itself, the best of any Claude models so far. Two of three outside groups Anthropic asked to review Claude found it rarely refused requests but sometimes added caveats or counterarguments nobody asked for, such as putting an opposing view into an advocacy email. A third group, the Foundation for American Innovation, found that an older model, Claude Sonnet 4.5, was more likely to soften arguments for right-leaning positions. Anthropic says it is working to make Claude add caveats more evenly across the political spectrum.
Anthropic has not said how well its new filters catch campaigns split into small steps, or how it will check that Claude's state-by-state voting answers stay correct as deadlines and polling places change. Anthropic says the voting answers are coming to U.S. users of Claude.ai; it has not said whether they will reach Claude's other apps.