Anthropic readies Claude for the 2026 US midterms: 0.01% of chats ask whom to vote for, official voting info and live results in answers, new classifiers for split-up influence campaigns

Why it matters

【Key points】

On Oct 14 (US time) Anthropic detailed Claude's election safeguards for the 2026 US midterms and shared usage data:

1. Usage: in ~250,000 sampled US Claude.ai chats (Sep 21–27), ~5.75% touched politics or policy and ~0.51% elections or voting; ~0.01% asked whom to vote for and ~0.002% broke election rules.

2. Official info: later this month Claude.ai will pull state-specific voting info from nonpartisan Democracy Works into answers, and show the latest official results on and after election night.

3. Bias: Anthropic says Opus, Sonnet and Haiku 5.5 score 99% on its even-handedness eval. Three outside reviews found mostly substantive answers but unsolicited caveats, and Sonnet 4.5 softening right-leaning arguments more often.

4. Misuse: red-teaming with ISD and the McCain Institute (US, Israel, Brazil) found Claude refuses single harmful requests but can miss a campaign split into harmless-looking tasks. New classifiers target multi-step misuse such as fake accounts and news sites.

5. Enforcement: accounts of a foreign fake news site pushing Russian, Iranian and North Korean narratives, and of ad-revenue fake news operators, were banned.

【Verifiable】

https://www.anthropic.com/news/2026-us-midterms

【Our read】

1. The key finding is the split-up attack: per-message refusal works, but a campaign broken into research, site building, drafting and distribution slips through. That needs cross-request detection, the same problem as multi-step agent misuse.

2. With only 0.01% asking whom to vote for, people use AI to research, not decide. Routing where-to-vote and results questions to official data is safer than model memory.

3. The 99% score is in-house while outside reviewers found uneven hedging; third-party retests on real chats matter more.