Preparing for the 2026 US midterms

What it claims

Anthropic describes its election safeguards for Claude ahead of the 2026 US midterms. The post is mainly useful for its named failure modes and the red-team method that exposed them. A one-week sample of 250,000 US Claude.ai conversations puts politics, policy and social issues at about 5.75%. Elections or voting specifically account for about 0.51%, requests for whom to vote for 0.01%, and policy-violating election requests about 0.002%. Anthropic says Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 score 99% on its internal political evenhandedness evaluation. It also says these models decline fewer political requests than earlier ones.

How it works

Voting logistics on Claude.ai are moving from a link to an in-answer feature. A banner linking to TurboVote, run by Democracy Works, already appears for registration, polling-place and election-date questions. Claude.ai will now draw official, state-specific voting information from Democracy Works into its answers for US users. That covers questions such as where to vote and what ID to bring. Claude will also show official vote counts on election night and in the days after.

Misuse detection uses automated classifiers. Flagged activity goes to threat intelligence and enforcement teams, who ban accounts once a violation is confirmed. New classifiers target multi-step misuse: fake social media accounts, fake political news sites, and financially motivated election schemes. Anthropic also added detection for foreign state actors, cyber protections for election infrastructure and campaigns, and country-specific safeguards for the US, Israel and Brazil.

Evidence

The Institute for Strategic Dialogue and the McCain Institute red-teamed Claude in the US, Israel and Brazil. They started with direct requests. They then split full influence campaigns into harmless-looking subtasks: audience research, building news sites, drafting content and planning distribution. Claude reliably refused single harmful requests. It could miss a malicious campaign spread across many benign-looking requests.

Three outside groups reviewed political responses. The Future of Free Speech, a Vanderbilt think tank, tested hundreds of contentious requests. The Foundation for American Innovation used matched left and right prompts, from mainstream to extreme. The Collective Intelligence Project built an evaluation from a three-stage process with more than 2,000 people in the US, UK and India. Most responses in the first two reviews were rated strong. No scores are given.

Enforcement cases include foreign-based operators of a fake news site. They used Claude to write US midterm coverage favouring Russian, Iranian and North Korean narratives. Other actors produced fabricated news aimed at American audiences to earn ad revenue. Anthropic banned the accounts in both cases.

Limits

The reviewers found Claude rarely refused requests. However, it sometimes added unsolicited caveats or counterarguments, such as an opposing view inserted into an advocacy email. The Foundation for American Innovation found Claude Sonnet 4.5 hedged unevenly and was more likely to soften arguments for right-leaning positions. Anthropic found answers to yes-or-no questions and requests for short answers were less balanced, and says it has improved both. The 99% evenhandedness figure comes from Anthropic's own evaluation, which so far covers policy issues. Extending it to candidates and voting questions is described as future work. The usage data covers one week of consumer Claude.ai traffic in the US and excludes API use. Prohibited campaign uses include fake profiles posing as real people, mass-produced comments to inflate engagement, and content for news sites that falsely claim independence.