Anthropic said Wednesday it will add official U.S. voting information and election results from nonprofit Democracy Works to Claude ahead of the Nov. 3 midterm elections. Later this month, Claude.ai users in the United States will be able to ask where to vote or what identification to bring and receive state-specific answers based on public information Democracy Works collects from state and local election sources. On election night and in the days afterward, Claude will also show official vote counts and results, Anthropic said.

Democracy Works first partnered with Anthropic in 2024 to prevent misuse of AI in political contexts. In April 2026, Anthropic began displaying the nonprofit’s nonpartisan TurboVote tool in a banner when users asked Claude about voter registration, polling locations or dates. Anthropic said its goal is to make Claude useful for political conversations while preventing election-related abuse.

Anthropic reported that politics, public policy or social issues appeared in about 5.75 percent of U.S. Claude.ai chats in a review of 250,000 conversations from late September. Elections or voting accounted for about 0.51 percent, requests for advice on whom or what to support for 0.01 percent, and policy violations for roughly 0.002 percent. The company said most political requests were informational, such as asking candidates’ positions, weighing policy arguments or drafting documents, and it expects political queries to rise as Election Day approaches.

Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 scored 99 percent on Anthropic’s political evenhandedness evaluation, the highest result to date, the company said. Independent reviews by the Future of Free Speech, the Foundation for American Innovation and the Collective Intelligence Project found the models rarely refused political requests but sometimes added unsolicited caveats. Tests by the Institute for Strategic Dialogue and the McCain Institute found Claude stopped single, clearly harmful requests but could miss coordinated activity split into smaller, harmless-looking steps. Anthropic said it added classifiers to detect novel, multi-step misuse, including attempts to build fake news sites. It said it will continue testing political evenhandedness and misuse as the election cycle intensifies.

Anthropic said Wednesday it will add official U.S. voting information and election results from nonprofit Democracy Works to Claude ahead of the Nov. 3 midterm elections. Later this month, Claude.ai users in the United States will be able to ask where to vote or what identification to bring and receive state-specific answers based on public information Democracy Works collects from state and local election sources. On election night and in the days afterward, Claude will also show official vote counts and results, Anthropic said.

Democracy Works first partnered with Anthropic in 2024 to prevent misuse of AI in political contexts. In April 2026, Anthropic began displaying the nonprofit’s nonpartisan TurboVote tool in a banner when users asked Claude about voter registration, polling locations or dates. Anthropic said its goal is to make Claude useful for political conversations while preventing election-related abuse.

Anthropic reported that politics, public policy or social issues appeared in about 5.75 percent of U.S. Claude.ai chats in a review of 250,000 conversations from late September. Elections or voting accounted for about 0.51 percent, requests for advice on whom or what to support for 0.01 percent, and policy violations for roughly 0.002 percent. The company said most political requests were informational, such as asking candidates’ positions, weighing policy arguments or drafting documents, and it expects political queries to rise as Election Day approaches.

Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 scored 99 percent on Anthropic’s political evenhandedness evaluation, the highest result to date, the company said. Independent reviews by the Future of Free Speech, the Foundation for American Innovation and the Collective Intelligence Project found the models rarely refused political requests but sometimes added unsolicited caveats. Tests by the Institute for Strategic Dialogue and the McCain Institute found Claude stopped single, clearly harmful requests but could miss coordinated activity split into smaller, harmless-looking steps. Anthropic said it added classifiers to detect novel, multi-step misuse, including attempts to build fake news sites. It said it will continue testing political evenhandedness and misuse as the election cycle intensifies.