A voting booth in Cuyahoga County, Ohio, during the state's primary election on March 15, 2016. Illustrative. Image: Tim Evanson / Wikimedia Commons, CC BY-SA 2.0, cropped
.jpg){kind=link}
Barely anyone asks Claude who to vote for. In a week-long sample of 250,000 US conversations on Claude.ai, requests for a voting recommendation came up in about 0.01% of chats, Anthropic said on Wednesday as it set out how it is preparing Claude for the November 3 US midterms.
The post pairs that usage data with a set of new safeguards: official, state-specific voting answers inside Claude, live results on election night, fresh political-bias testing by outside groups and new classifiers aimed at influence campaigns that hide in harmless-looking requests.
What people actually ask Claude about politics #
Anthropic ran the sample from September 21 to 27, 2026, through Anthropic Insights, its privacy-preserving analysis tool, and had Claude classify the conversations. Its figures:
- Politics, public policy or social issues: about 5.75% of all US conversations on Claude.ai, from following the news and research to learning about candidates and dealing with government services.
- Elections or voting specifically: about 0.51%.
- Whom or what to vote for: about 0.01%.
- Requests breaking its election integrity rules: about 0.002%, from a separate review of 250,000 conversations.
Most political requests are informational, Anthropic says: people use Claude to make sense of geopolitical events, research candidates, draft documents, help with coursework and weigh arguments for and against policies. It says users more often ask questions that help them reach their own conclusions, such as where candidates stand on an issue, and it expects the share of election queries to rise as Election Day nears.
State-by-state voting answers and election-night results #
Claude.ai already shows a banner linking to TurboVote, a nonpartisan tool from the nonprofit Democracy Works, when people ask about voter registration, polling places or election dates. Later this month, Claude will start drawing official voting information from Democracy Works into its answers for US users, so someone asking where to vote or what ID to bring gets a reply specific to their state without leaving the chat. That information comes from state and local election sources.
Claude will also show up-to-date official vote counts on election night and in the days that follow. Both features are for Claude.ai users in the US only.
Outside testers found Claude went softer on the right #
Anthropic says Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 are its most even-handed models yet, each scoring 99% on its own political even-handedness evaluation, and that they decline fewer political requests than before. Those are the company's own numbers.
It also asked three independent groups to test Claude. The Future of Free Speech, a think tank at Vanderbilt University, ran hundreds of contentious requests; the Foundation for American Innovation compared answers to matched prompts from the left and right; and the Collective Intelligence Project built an evaluation with more than 2,000 people in the US, UK and India. The first two rated most responses as strong.
The flaws are the more interesting part. The Future of Free Speech and the Collective Intelligence Project found Claude sometimes added unsolicited caveats and counterarguments, such as slipping an opposing view into an advocacy email. The Foundation for American Innovation found that Claude Sonnet 4.5 "didn’t always hedge evenly, and was more likely to soften arguments for right-leaning positions." Anthropic says it has improved answers to yes-or-no questions and short-answer requests, where balance slipped most, and is widening its bias tests to cover candidates and voting questions.
Fake news sites and campaigns split into small tasks #
The Institute for Strategic Dialogue and the McCain Institute pressure-tested Claude in the US, Israel and Brazil, all of which hold major national elections this year. They found Claude "reliably refused a single harmful request, but could miss a malicious campaign" when it was broken into many requests that each looked harmless, such as audience research, building a news site and planning distribution.
Anthropic says it has since added classifiers for multi-step misuse, including fake social media accounts and political news sites, plus better detection of foreign state actors. It describes one case in which foreign-based operators of a fake news site used Claude to produce US midterm coverage favouring Russian, Iranian and North Korean narratives, and others where people churned out fabricated news to earn advertising revenue. The accounts were banned. The company tightened the election wording in its usage policy last week and has published a guide for campaigns on what is allowed.
Why it matters #
AI money is already shaping these midterms, with rival super PACs tied to OpenAI and Anthropic figures spending heavily. Anthropic's own data says voters mostly use Claude to look things up, not to be told how to vote, but its testers found the real gap is campaigns sliced into innocent-looking pieces, and that is the weakness to watch on November 3.
Sources: Anthropic: Preparing for the 2026 US midterms; Anthropic: guide for political campaigns (PDF)