Skip to content
AI.info

The Pulse

Anthropic Finds Elections in 0.51% of Sampled US Claude.ai Consumer Conversations

A sample of 250,000 US Claude.ai consumer conversations found that 0.51% were about elections or voting, while 5.75% touched on politics, public policy or social issues. Claude will add state-specific voting information later in October and show official election results on election night and in the days afterward, Anthropic says.

Anthropic Finds Elections in 0.51% of Sampled US Claude.ai Consumer Conversations

Only 0.51% of sampled US conversations on Claude.ai concerned elections or voting, according to a new Anthropic analysis of usage ahead of the 2026 midterms. The company’s sample, drawn from 250,000 consumer conversations between September 21 and 27, found that 5.75% touched more broadly on politics, public policy or social issues. Requests asking Claude whom or what to vote for appeared in about 0.01% of conversations. Anthropic published the figures on October 14 as part of an update on its election safeguards.

What the 250,000-chat sample shows

Anthropic used its Anthropic Insights analysis tool to classify a random sample of US Claude.ai conversations in a privacy-preserving way. The figures cover the company’s consumer app, not all Claude use across its products. Political conversations included news, candidate research, policy questions, coursework and government services, the company said.

Requests that violated Anthropic’s election integrity policies appeared in about 0.002% of conversations. That figure came from a separate analysis of 250,000 conversations reviewed against those policies, rather than from the September 21–27 sample used for the broader usage breakdown, the company says. Anthropic’s chart also presents the upper end of a 95% confidence interval for requests seeking a voting recommendation, because those requests were rare.

Claude will add state-specific voting details

Anthropic plans to add official voting information from Democracy Works to Claude.ai responses for US users later in October. Someone asking where to vote or what identification to bring should receive information specific to their state, based on public information Democracy Works gathers from state and local election sources. Claude will continue showing a banner that directs people asking about registration, polling places or election dates to TurboVote, Democracy Works’ nonpartisan resource.

Claude will also show up-to-date election results on election night and in the days afterward. People asking for vote counts or results will see the latest official numbers in their conversations. Voting information is due later in October; official results will appear on election night and in the days that follow.

Outside evaluations found uneven answers

Anthropic’s newest models—Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5—scored 99% on the company’s political even-handedness evaluation, Anthropic says. Three outside groups reviewed related questions using different methods: The Future of Free Speech at Vanderbilt University tested requests on contentious topics, the Foundation for American Innovation compared matched prompts from the left and right, and the Collective Intelligence Project developed an evaluation with input from more than 2,000 people in the US, UK and India.

The results also identified weaknesses. The Future of Free Speech and the Collective Intelligence Project found that Claude sometimes inserted unsolicited caveats or counterarguments, Anthropic said. The Foundation for American Innovation found that Claude Sonnet 4.5 was more likely to soften arguments for right-leaning positions. Anthropic has adjusted responses to direct yes-or-no questions and requests for shorter answers, and plans to expand testing to cover candidates and voting questions alongside policy issues.

Anthropic describes fake-news and financial schemes

Anthropic worked with the Institute for Strategic Dialogue and the McCain Institute to test its defenses against deceptive campaigning, misinformation and other election abuse in the US, Israel and Brazil. In those tests, Claude refused direct harmful requests but could miss a coordinated operation when its component tasks looked harmless on their own. Anthropic added classifiers intended to detect multi-step activity, including attempts to build fake social-media accounts or political news sites.

Anthropic also described investigations involving the US midterms. Operators of a fake news site based abroad used Claude to produce coverage favoring Russian, Iranian and North Korean narratives, the company says; other users generated fabricated news to attract US audiences and advertising revenue, it also says. Anthropic banned the accounts involved and improved detection. The company plans to publish a broader review of Claude use during the election cycle in the coming months.

Sources

  1. Original source Anthropic anthropic.com

Companies in this story

Related stories

Explore

More articles