Preparing for the 2026 US midterms

Summary

Anthropic describes how Claude is used for election topics, the voter-information features it is adding before the 2026 US midterms, its political-bias testing, and the election misuse it has detected and banned. It updates the company's earlier election safeguards work.

Key points

• Usage, from 250,000 sampled US Claude.ai conversations (2026-09-21 to 09-27): about 5.75% touch politics or social issues, 0.51% elections or voting, 0.01% ask whom to vote for, and about 0.002% violate election-integrity policies.
• Since April a TurboVote banner (Democracy Works) appears for registration and polling questions. Later in October 2026, US Claude.ai users will get state-specific official voting information inside the chat, and official vote counts on election night.
• Opus 5.5, Sonnet 5.5, and Haiku 5.5 score 99% on Anthropic's political evenhandedness evaluation, its best so far; they also refuse fewer political requests.
• External reviewers (Future of Free Speech, Foundation for American Innovation, Collective Intelligence Project) rated most responses strong. Weaknesses: unsolicited caveats and counterarguments, and Sonnet 4.5 hedging unevenly (softening right-leaning arguments more).
• Red teaming by the Institute for Strategic Dialogue and McCain Institute (US, Israel, Brazil) found Claude refuses single harmful requests reliably but can miss campaigns split into harmless-looking subtasks. New classifiers target this multi-step misuse.
• Banned operations: a foreign fake-news site pushing Russian, Iranian, and North Korean narratives, and ad-revenue schemes producing fabricated US news.

Why it matters

It is a concrete case of the "decomposed task" weakness in refusal-based safety: single requests are caught, but distributed campaigns need account-level detection. It also gives rare usage figures for how often people ask a chatbot about elections.