Preparing for the 2026 US midterms

Anthropic outlined how it has strengthened its election safeguards ahead of the 2026 US midterms, building on earlier work. Analysis of 250,000 Claude conversations found that about 5.75% of US conversations touch on politics, 0.51% concern elections or voting, and only 0.01% ask whom or what to vote for, while policy-violating requests are about 0.002%. To connect users with trusted information, Claude.ai displays a TurboVote banner for voter registration, polling place, and election date questions, and will soon pull official voting information from Democracy Works directly into state-specific answers, plus show up-to-date election results on election night. On political bias, Anthropic says Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 are its most evenhanded models yet, scoring 99% on its evenhandedness evaluation, though independent reviews by The Future of Free Speech, the Foundation for American Innovation, and the Collective Intelligence Project found remaining issues such as unsolicited caveats and uneven hedging. The company updated its Usage Policy, published a campaign guide, and added classifiers to catch multi-step and novel misuse, including fake news sites and financially motivated election schemes, after pressure-testing with the Institute for Strategic Dialogue and the McCain Institute in the US, Israel, and Brazil. Threat intelligence work uncovered foreign operators using Claude for fake US midterm news favoring Russian, Iranian, and North Korean narratives, and financially motivated fabricated content; those accounts were banned. Anthropic says it will continue investing in evenhandedness, misuse prevention, and transparency.

(Source：Anthropic)