Anthropic has outlined a set of strengthened election safeguards for its Claude models ahead of the 2026 US midterms, building on an update it published earlier this year on measuring political bias, detecting misuse, and pointing users toward accurate voting information.
To understand how people actually use Claude in political contexts, the company analyzed 250,000 conversations over a recent one-week period using Anthropic Insights. It found that roughly 5.75% of US conversations on Claude.ai touched broadly on politics, public policy, or social issues, while about 0.51% were specifically about elections or voting. Requests asking Claude who or what to vote for appeared in only about 0.01% of conversations, and policy-violating requests in about 0.002%.
On the information side, Anthropic said Claude.ai will continue to display a banner linking to TurboVote, a nonpartisan resource from Democracy Works, for questions about voter registration, polling places, or election dates. Later this month, Claude will begin pulling official voting information from Democracy Works directly into responses for US users, offering state-specific answers on where to vote or what ID to bring. Claude will also show up-to-date official election results on election night and in the days after.
Anthropic said its latest models — Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 — are its most even-handed to date, scoring 99% on its political even-handedness evaluation. Three independent groups reviewed Claude's handling of free expression: the Future of Free Speech, the Foundation for American Innovation, and the Collective Intelligence Project. Most responses were rated as strong, though the reviews also flagged occasional unsolicited caveats and uneven hedging, including a finding that Claude Sonnet 4.5 was more likely to soften arguments for right-leaning positions. Anthropic said it is improving how Claude handles yes-or-no questions and when it adds counterarguments.
On enforcement, Anthropic partnered with the Institute for Strategic Dialogue and the McCain Institute to pressure-test protections in the US, Israel, and Brazil. Analysts found that Claude reliably refused single harmful requests but could miss malicious campaigns split across many individually harmless tasks. In response, Anthropic added classifiers to detect novel and multi-step misuse, enhanced detection of foreign state-actor activity, strengthened cyber protections for election infrastructure, and built localized safeguards for the three countries, all of which hold major national elections this year.