Anthropic says few US Claude chats ask whom to vote for

In a sample of about 250,000 U.S. chats, at most about 0.01% asked whom or what to vote for and about 0.51% were about elections or voting, Anthropic said.

Anthropic said Wednesday that very few U.S. users ask its Claude chatbot whom or what to vote for, according to a company post laying out its safeguards for the midterms.

The company said it studied a random sample of about 250,000 U.S. Claude.ai consumer conversations from Sept. 21 to 27.

The numbers

All figures are Anthropic’s own, from its own sample.

About 5.75% of U.S. Claude.ai conversations touched broadly on politics, public policy or social issues, the company said.
About 0.51% were specifically about elections or voting.
At most about 0.01% asked Claude whom or what to vote for. Anthropic said such requests are so rare that it reports the upper end of a 95% confidence interval.
About 0.002% violated its election integrity policies. Anthropic said that figure came from a separate analysis of 250,000 conversations.

Anthropic said it expects the share of political and election questions to increase as Election Day nears.

The rule in English

Anthropic’s Usage Policy bars using Claude to deceive voters or disrupt elections. That includes spreading false information about candidates or how to vote, impersonating election officials and trying to suppress turnout, the company said.

Who wrote it

Anthropic sets the rules. It said it updates the policy every year. This year’s update clarifies what the election rules cover and adds examples.

Who it binds

The rules apply to anyone using Claude, including campaigns. Anthropic said campaigns may not use Claude to create fake social media profiles posing as real people, mass-produce comments to inflate engagement or write for news sites that falsely claim to be independent.

Who it exempts

Anthropic said legitimate campaign work is allowed. Its examples include drafting press statements and fundraising emails, prepping for debates, analyzing turnout data and building voter outreach materials. The company published a guide for campaigns.

The safeguards

Anthropic said it:

Shows a banner linking to TurboVote, a nonpartisan resource from Democracy Works, when users ask about registration, polling places or election dates.
Will start drawing official voting information from Democracy Works into Claude.ai answers for U.S. users later this month, so a question about where to vote or what ID to bring gets a state-specific answer.
Will show up-to-date official results on election night and in the days after.
Tests models for political bias before every launch. It said its newest models score 99% on its political evenhandedness evaluation.
Asked three outside groups to review how Claude handles free expression. Two of them found Claude sometimes added unsolicited caveats and counterarguments, and one found an earlier model was more likely to soften arguments for right-leaning positions, Anthropic said. The company said it is addressing both.
Uses automated classifiers to detect election abuse. Testing with the Institute for Strategic Dialogue and the McCain Institute found Claude could miss a campaign split into many harmless-looking requests, Anthropic said. It said it added classifiers for multistep misuse, fake accounts and fake news sites.
Strengthened detection of misuse by foreign state actors and cyber protections for election infrastructure and campaigns.
Has teams review flagged activity and ban accounts when they confirm a violation.

Anthropic said it banned accounts tied to a foreign-based fake news site that used Claude to produce midterm coverage favoring Russian, Iranian and North Korean narratives. It said it also banned accounts using Claude to fabricate news for ad revenue.