AI and US Midterms: Anthropic’s Approach to Bias and Integrity

Insights on Political Engagement Through AI

Very few people asked Claude whom they should vote for. These questions made up just 0.01 percent of the conversations analyzed in the United States.

According to Anthropic, about 0.002 percent of these conversations included requests that went against the company's rules for protecting election integrity.

A detailed review of political discussions showed how people use the chatbot. About 35.8 percent of conversations focused on writing and academic help, while 29.7 percent involved academic and professional research.

Questions about politics and world affairs made up 10.7 percent of conversations. Political fact-checking and discussions of arguments accounted for a similar share, at 10.6 percent.

Other activities included learning about elections and voting, studying economic and financial issues linked to world affairs, and examining political or philosophical ideas.

These findings suggest that people use AI for political topics in ways that go beyond simply asking which candidate to choose.

People also use it to understand policy debates, explore issues affecting their communities, compare different arguments, and find useful information that helps them take part in civic life.

Evaluating Political Bias in AI Models

Anthropic says it regularly tests its models for political bias before releasing them. It aims to make sure that answers about different viewpoints offer a similar level of detail, consistency, and careful analysis.

In its latest review ahead of the midterm elections, the company reported that Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 scored 99 percent on a test of political impartiality.

The company also asked three independent organizations to examine how Claude handles political discussions and other controversial topics.

These groups were The Future of Free Speech, an independent think tank connected to Vanderbilt University; the Foundation for American Innovation; and the Collective Intelligence Project. The Collective Intelligence Project surveyed more than 2,000 people in the United States, the United Kingdom, and India while developing a framework for evaluating the models.

The evaluations found that Claude responded meaningfully to users' questions, but they also identified some problems.

For example, Claude sometimes added extra warnings, qualifications, or opposing arguments that users had not asked for. This sometimes happened even in emails supporting a particular position.

Another evaluation found that an earlier version, Claude Sonnet 4.5, tended to weaken arguments that supported right-leaning views.

In response, Anthropic has worked to make Claude follow direct yes-or-no questions and short prompts more closely. It has also adjusted when the chatbot adds qualifications or presents opposing views.

The company is also expanding its tests to cover real-world conversations, questions about candidates, and election information, rather than focusing mainly on broad policy debates.

Accessing Official Voting Information via Claude

Another part of Anthropic's plan is to help users find trustworthy election information without leaving the chatbot.

The company said Claude would soon provide official U.S. voting information from Democracy Works, a nonprofit organization that collects publicly available information from state and local election authorities.

This feature is meant to answer practical questions about voter registration, polling locations, and the identification needed to vote in each user's state.

Anthropic also said it plans to keep showing a banner that directs users who ask about voting rules or election dates to TurboVote, a nonpartisan service run by Democracy Works.

The organization also plans to provide updated election results on election night and in the following days, allowing users to get the latest official information through conversations with Claude.

According to Reuters, as of October 14, Anthropic was preparing to launch official voting information and election results before the upcoming November midterm elections.

Measures to Prevent Election Manipulation

Anthropic has also put safeguards in place to stop people from using Claude to carry out deceptive political activities.

The company's usage rules clearly ban actions such as spreading false information about candidates or voting procedures, pretending to be election officials, stopping people from voting, and creating fake political activity to mislead others.

However, legitimate campaign work is allowed. This includes writing press statements, preparing for debates, studying voter turnout data, and creating materials to reach voters.

By contrast, making fake social media accounts, producing large numbers of comments to create the false impression of public support, and creating content for websites falsely presented as independent news outlets are strictly forbidden.

Anthropic uses automated systems called classifiers to identify possible election-related misuse.

Its threat intelligence and enforcement teams review flagged activity and can block accounts that break the company's rules.

The company has also worked with the Institute for Strategic Dialogue and the McCain Institute to test its defenses against deceptive campaign activity and other election-related abuses in the United States, Israel, and Brazil.

Anthropic said these tests found that Claude consistently refused individual harmful requests. However, it sometimes had trouble recognizing coordinated harmful operations when people divided them into smaller tasks that seemed harmless on their own.

In response, the company added more classifiers to detect campaigns made up of several steps, including efforts to create fake social media accounts and false political news websites.

Anthropic also improved its ability to detect election-related misuse by foreign governments and strengthened cyber protections for election systems and political campaigns.

Combating Disinformation and Foreign Influence

Anthropic revealed that foreign operators behind a fake news outlet had used Claude to produce biased coverage of the U.S. midterm elections. The coverage supported stories and claims that aligned with Russia, Iran, and North Korea.

The company said it responded by banning the accounts involved and improving its systems for detecting this kind of activity.

There were also reports of people trying to make money from the election period. These actors used Claude to create fictional news stories and other content designed to attract American readers and earn money from advertising.

A separate report by AFP described a propaganda operation linked to a French advertising agency.

According to the report, the operation used Claude to produce large numbers of political articles for about 70 fake news websites in around 20 languages. It published at least 8,913 stories and targeted several countries, including the United States, Brazil, and France.

Meanwhile, The Associated Press reported that political advertisements created or improved with AI had started appearing during the 2026 U.S. midterm election cycle. Rules requiring these advertisements to disclose their use of AI were not being followed consistently.