What shipped: Claude's 2026 midterm election safeguards

Ahead of the 2026 US midterms, Anthropic strengthened how Claude handles political and election questions: live, state-specific voting info from Democracy Works, a 99% even-handedness score on Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5, and new classifiers that catch election misuse spread across many small requests. This lesson shows you how to use Claude for political research and campaign work the way it's designed to be used now.

Claude can now answer "where do I vote?" with your actual polling place. Later this month, Claude starts pulling official voting information — polling locations, required ID, election dates — from Democracy Works directly into its answer for US users. Previously you'd just get a banner linking out; now the answer is in the conversation. Claude will also surface up-to-date official results on election night and the days after.

Political conversations are common, but vote-recommendation requests are almost nonexistent. In a sample of 250,000 US Claude.ai conversations (Sept 21–27, 2026), 5.75% touched politics, public policy, or social issues, and 0.51% were specifically about elections or voting. Only 0.01% of conversations asked Claude who or what to vote for — and only 0.002% violated Anthropic's election integrity policy. Knowing this helps you predict what Claude will and won't do: it treats "help me understand this candidate's positions" very differently from "tell me who to vote for."

Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 score 99% on Anthropic's political even-handedness evaluation — the highest yet. That means when you ask Claude to lay out arguments on a contested policy issue, it should engage substantively rather than refuse, and it shouldn't tilt toward one side. Independent reviewers (The Future of Free Speech at Vanderbilt, the Foundation for American Innovation, and the Collective Intelligence Project, which surveyed 2,000+ people across the US, UK, and India) found a real gap worth knowing about: Claude sometimes adds unsolicited caveats or counterarguments even when you didn't ask for them — for example, slipping an opposing view into an email you asked it to write advocating for one side. If you want a one-sided persuasive draft, say so explicitly.

Claude now catches election misuse that's split across many small, innocent-looking requests. Working with the Institute for Strategic Dialogue and the McCain Institute, Anthropic found that Claude reliably refuses a single harmful request (like "write fake news favoring Candidate X") but could miss a coordinated campaign broken into separate steps — audience research, site-building, content drafting, distribution planning — each of which looks harmless alone. New classifiers now target this pattern, plus fake social-media accounts, fake news sites, and foreign state-actor influence operations. In one real case, a foreign-based fake news site used Claude to produce US midterm coverage favoring Russian, Iranian, and North Korean narratives; the accounts were banned.

Claude now gives you state-specific voting info and live election results directly in chat, scores 99% on even-handedness across Opus 5.5, Sonnet 5.5, and Haiku 5.5, and won't tell you who to vote for — but will help you compare candidates' positions if you ask. If you're doing campaign work, legitimate tasks (drafts, turnout analysis, debate prep) are fine; fake accounts, fake news sites, and engagement-inflation requests are not, even split across multiple prompts.