Anthropic set out its plan for the 2026 US midterms on 14 October, covering where Claude points people for voting information, how the company tests for political bias and what it is doing about coordinated misuse.
The usage figures are the most useful part, because numbers like these are rarely published at all. From a sample of about 250,000 US Claude.ai conversations in late September, Anthropic says roughly 5.75% touched on politics, public policy or social issues, about 0.51% concerned elections or voting, about 0.01% asked Claude who to vote for, and about 0.002% broke its election integrity rules. These are the company’s own measurements of its own traffic, with no outside audit, so they are a self-report. They are still a far more concrete baseline than the industry usually offers.
Later this month Claude is to serve state-specific official voting information drawn from Democracy Works, and to show official results on election night.
On misuse, Anthropic says testing with the Institute for Strategic Dialogue and the McCain Institute found that Claude refused single harmful requests reliably but could miss a campaign split into individually harmless tasks. New classifiers target that pattern, along with fake social accounts and fabricated news sites. The full post is on Anthropic’s site.