Anthropic said it has expanded safeguards for Claude ahead of the 2026 US midterms, adding official state-specific voting information, election-night results and tools to detect political abuse. The company also reported that a review of about 250,000 conversations found requests asking whom or what to vote for in about 0.01% of US Claude conversations, while requests violating its election integrity policies appeared in about 0.002%.

Anthropic Insights analyzed a random sample of about 250,000 US consumer conversations from September 21 to 27, 2026, using a privacy-preserving method. About 5.75% of US conversations on Claude.ai touched broadly on politics, public policy or social issues, and about 0.51% concerned elections or voting. Anthropic said requests for voting recommendations were rare; more commonly, people sought information about candidates, policies, government services, coursework and geopolitical events. The company expects the share of political and election-related queries to increase as Election Day approaches.

For users in the US, Claude will later this month begin incorporating voting information collected by Democracy Works from state and local election sources. Answers to questions about where to vote or what identification to bring will be tailored to the user's state. Claude will also show official vote counts and election results on election night and in the days afterward. Anthropic said that, as it shared in April, questions about registration, polling places or election dates will continue to trigger a banner linking to TurboVote, a nonpartisan resource from Democracy Works.

Anthropic said it tests every model launch for political bias. Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 scored 99% on its political evenhandedness evaluation, which the company described as its strongest result to date. The models also engage more readily in political discussions and decline fewer requests. Three independent groups reviewed Claude's handling of free expression: The Future of Free Speech tested hundreds of contentious requests; the Foundation for American Innovation compared matched prompts from the political left and right; and the Collective Intelligence Project surveyed more than 2,000 people in the US, UK and India to develop evaluation principles.

The reviews also found weaknesses. The Future of Free Speech and the Collective Intelligence Project found that Claude rarely refused requests but sometimes added unsolicited caveats or counterarguments, including opposing views in advocacy emails. The Foundation for American Innovation found that Claude Sonnet 4.5 did not always hedge evenly and was more likely to soften arguments for right-leaning positions. Anthropic said it has improved responses to direct yes-or-no questions and requests for shorter answers, and is refining how the models add caveats and how it tests for uneven hedging. It plans to broaden evaluations to reflect real conversations and include candidates and voting questions as well as policy issues.

Anthropic's Usage Policy bars using Claude to deceive voters or disrupt elections, including spreading false voting information, impersonating election officials and suppressing turnout. An update announced last week clarifies the rules and adds examples. A recently published guide explains permitted campaign tasks, such as drafting press statements and fundraising emails, preparing for debates, analyzing turnout data and making voter outreach materials. It also gives prohibited examples: fake social media profiles posing as real people, mass-produced comments intended to inflate campaign engagement, and content for pages falsely presented as independent news sites.

Anthropic said it partnered with the Institute for Strategic Dialogue and the McCain Institute to test safeguards in the US, Israel and Brazil. Researchers tried direct harmful requests and campaigns divided into smaller tasks, such as audience research, building news sites, drafting content and planning distribution. Claude reliably refused single harmful requests but sometimes missed malicious campaigns when their component requests appeared harmless. Anthropic said it added classifiers to detect novel and multi-step misuse, including fake accounts and political news sites, financially motivated election schemes, foreign state activity and threats to election infrastructure and campaigns. It also added safeguards for localized risks

Anthropic said it has expanded safeguards for Claude ahead of the 2026 US midterms, adding official state-specific voting information, election-night results and tools to detect political abuse. The company also reported that a review of about 250,000 conversations found requests asking whom or what to vote for in about 0.01% of US Claude conversations, while requests violating its election integrity policies appeared in about 0.002%.

Anthropic Insights analyzed a random sample of about 250,000 US consumer conversations from September 21 to 27, 2026, using a privacy-preserving method. About 5.75% of US conversations on Claude.ai touched broadly on politics, public policy or social issues, and about 0.51% concerned elections or voting. Anthropic said requests for voting recommendations were rare; more commonly, people sought information about candidates, policies, government services, coursework and geopolitical events. The company expects the share of political and election-related queries to increase as Election Day approaches.

For users in the US, Claude will later this month begin incorporating voting information collected by Democracy Works from state and local election sources. Answers to questions about where to vote or what identification to bring will be tailored to the user's state. Claude will also show official vote counts and election results on election night and in the days afterward. Anthropic said that, as it shared in April, questions about registration, polling places or election dates will continue to trigger a banner linking to TurboVote, a nonpartisan resource from Democracy Works.

Anthropic said it tests every model launch for political bias. Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 scored 99% on its political evenhandedness evaluation, which the company described as its strongest result to date. The models also engage more readily in political discussions and decline fewer requests. Three independent groups reviewed Claude's handling of free expression: The Future of Free Speech tested hundreds of contentious requests; the Foundation for American Innovation compared matched prompts from the political left and right; and the Collective Intelligence Project surveyed more than 2,000 people in the US, UK and India to develop evaluation principles.

The reviews also found weaknesses. The Future of Free Speech and the Collective Intelligence Project found that Claude rarely refused requests but sometimes added unsolicited caveats or counterarguments, including opposing views in advocacy emails. The Foundation for American Innovation found that Claude Sonnet 4.5 did not always hedge evenly and was more likely to soften arguments for right-leaning positions. Anthropic said it has improved responses to direct yes-or-no questions and requests for shorter answers, and is refining how the models add caveats and how it tests for uneven hedging. It plans to broaden evaluations to reflect real conversations and include candidates and voting questions as well as policy issues.

Anthropic's Usage Policy bars using Claude to deceive voters or disrupt elections, including spreading false voting information, impersonating election officials and suppressing turnout. An update announced last week clarifies the rules and adds examples. A recently published guide explains permitted campaign tasks, such as drafting press statements and fundraising emails, preparing for debates, analyzing turnout data and making voter outreach materials. It also gives prohibited examples: fake social media profiles posing as real people, mass-produced comments intended to inflate campaign engagement, and content for pages falsely presented as independent news sites.

Anthropic said it partnered with the Institute for Strategic Dialogue and the McCain Institute to test safeguards in the US, Israel and Brazil. Researchers tried direct harmful requests and campaigns divided into smaller tasks, such as audience research, building news sites, drafting content and planning distribution. Claude reliably refused single harmful requests but sometimes missed malicious campaigns when their component requests appeared harmless. Anthropic said it added classifiers to detect novel and multi-step misuse, including fake accounts and political news sites, financially motivated election schemes, foreign state activity and threats to election infrastructure and campaigns. It also added safeguards for localized risks