Anthropic is expanding election-related safeguards for Claude ahead of the 2026 U.S. midterm elections, introducing additional voting-information features, political-bias evaluations, and systems designed to detect coordinated attempts to misuse its AI models. The company said the work builds on safeguards previously introduced to measure political bias, detect election-related abuse, and connect users with accurate voting information.
Anthropic analyzed a random sample of approximately 250,000 U.S. Claude.ai consumer conversations from September 21 through September 27, 2026. Conversations broadly involving politics, public policy, or social issues represented about 5.75% of U.S. conversations, while approximately 0.51% were specifically related to elections or voting. Requests asking Claude whom or what to vote for appeared in roughly 0.01% of conversations, while requests violating Anthropic’s election-integrity policies represented about 0.002%.
The company said most political conversations are informational rather than requests for voting recommendations. Users turn to Claude to follow geopolitical developments, research candidates, draft documents, complete coursework, and evaluate arguments for and against public policies.
Anthropic is also expanding access to official election information inside Claude. Users asking about voter registration, polling locations, or election dates will continue to see links to TurboVote, the nonpartisan service operated by Democracy Works. Beginning later in October, Claude is expected to incorporate official Democracy Works voting information directly into U.S. responses, allowing users to receive state-specific information such as polling locations and identification requirements without leaving the conversation.
Claude will additionally begin displaying current election results on election night and in the days following the vote. Anthropic said users asking about vote counts or election outcomes will be shown the latest official numbers directly within their conversations.
Anthropic said Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 each scored 99% on its political-evenhandedness evaluation, making them the company’s most evenhanded models to date. The company also said the newer models engage more readily in political conversations while declining fewer legitimate requests.
To further evaluate political behavior, Anthropic worked with independent groups including The Future of Free Speech at Vanderbilt University, the Foundation for American Innovation, and the Collective Intelligence Project. Those evaluations generally found Claude substantively engaged with political requests, although researchers also identified areas for improvement, including unsolicited caveats, counterarguments, and uneven hedging. One evaluation found that Claude Sonnet 4.5 was more likely to soften arguments supporting right-leaning positions. Anthropic said it is adjusting how Claude responds to direct questions, applies caveats, and handles matched political prompts across ideological perspectives.
Anthropic’s Usage Policy prohibits using Claude to deceive voters or disrupt elections, including spreading false information about candidates or voting procedures, impersonating election officials, suppressing turnout, creating fake social media identities, manufacturing artificial campaign engagement, or producing content for sites falsely presented as independent news organizations. The company said legitimate campaign uses can include drafting press statements and fundraising emails, preparing for debates, analyzing turnout data, and developing voter outreach materials.
The company has also strengthened automated classifiers designed to detect election-related abuse after working with the Institute for Strategic Dialogue and the McCain Institute to test its defenses. Researchers found Claude reliably refused individual harmful requests but could sometimes miss coordinated campaigns when malicious activity was divided among numerous apparently harmless tasks. Anthropic subsequently added classifiers focused on multi-step abuse, fake political social media accounts and news sites, financially motivated election schemes, foreign state actors, and attacks against election infrastructure and campaigns.
Anthropic said its threat-intelligence teams have observed political, state-aligned, and commercial actors integrating AI throughout influence operations. In one case, foreign-based operators of a fake news website used Claude to generate U.S. midterm coverage favoring Russian, Iranian, and North Korean narratives. The company also identified actors using fabricated election news to attract American audiences and generate advertising revenue. Anthropic said it banned the accounts involved and strengthened its systems for detecting similar activity.
