Anthropic · Thursday, October 15, 2026
Anthropic says a small share of Claude chats are about voting
The File desk · Oct 15, 2026, 11:00 AM UTC
Status
Confirmed as Anthropic’s account, posted Oct. 14, 2026. The percentages are the company’s counts of its own conversations and its own evenhandedness test. This desk did not see the underlying chats.
- Confirmed
Anthropic said a random sample of 250,000 U.S. Claude.ai consumer conversations from Sept. 21 to Sept. 27, 2026, found about 5.75 percent touching politics, public policy, or social issues.
- Confirmed
The company said about 0.51 percent of all conversations were specifically about elections or voting, and about 0.01 percent asked whom or what to vote for.
- Confirmed
A separate sample of 250,000 conversations found policy violations in about 0.002 percent. Anthropic said Opus 5.5, Sonnet 5.5, and Haiku 5.5 scored 99 percent on its political evenhandedness evaluation.
Anthropic said Wednesday that, in a random sample of 250,000 U.S. Claude.ai consumer conversations from Sept. 21 to Sept. 27, about 5.75 percent touched politics, public policy, or social issues. About 0.51 percent of all conversations were about elections or voting. About 0.01 percent asked whom or what to vote for. Those shares are Anthropic’s. Later this month, the company said, Claude.ai will show U.S. users voting information from Democracy Works and election-night results. A TurboVote banner was already on the product.
The post is dated Oct. 14, 2026, and titled as a preparation note for the 2026 U.S. midterms. The sample is consumer chats on Claude.ai in the United States, not a poll of voters.
Anthropic said a separate analysis of 250,000 conversations found policy violations in about 0.002 percent of them. It said Opus 5.5, Sonnet 5.5, and Haiku 5.5 scored 99 percent on the company’s political evenhandedness evaluation. That is Anthropic’s test.
The company said testing with the Institute for Strategic Dialogue and the McCain Institute found that the model refused a single harmful request but could miss a campaign that was split across steps that each looked harmless. That finding is as Anthropic describes the outside test.
On the product, Anthropic said a TurboVote banner was already in place. Later this month, it said, Claude.ai will draw Democracy Works voting information for U.S. users and show election-night results. This desk did not see that later feature live.
What is still unknown or disputed
- The conversation samples were not opened. The percentages are Anthropic’s.
- The 99 percent evenhandedness score is the company’s evaluation, not an outside audit this desk ran.
- Democracy Works information and election-night results are described as coming later this month. They were not a finished feature on the page beyond the existing TurboVote banner.
Primary sources
Every claim in this story is drawn from the documents below. If a fetch failed, that is recorded on the card.
Source 1
Preparing for the 2026 US midterms
Anthropic · October 14, 2026
A random sample of 250,000 U.S. Claude.ai consumer conversations from September 21 to 27, 2026. About 5.75 percent touched politics, public policy, or social issues. About 0.51 percent were about elections or voting. About 0.01 percent asked whom or what to vote for. Policy violations about 0.002 percent in a separate sample of 250,000.
https://www.anthropic.com/news/2026-us-midterms