Anthropic strengthens Claude ahead of U.S. elections

Anthropic is preparing Claude for the United States midterm elections with new official sources, controls against political abuse and systems designed to detect coordinated disinformation campaigns.

The company analyzed 250,000 conversations involving Claude in the United States over one week to understand how people use the assistant for election-related matters. The results show that most users are not asking it to choose for them.

What users ask Claude

Conversations about politics, public policy or social issues account for around 5.75% of all conversations on Claude.ai in the United States. Queries specifically related to elections or voting account for 0.51%.

According to Anthropic's analysis, requests for recommendations about whom to vote for or which option to support appeared in just 0.01% of conversations. Requests that violated election integrity rules were even less common, at approximately 0.002%.

In practice, users are more likely to turn to Claude to:

• Understand a candidate's position on an issue.
• Follow current events and research public policy.
• Find information about government services.
• Check how to register to vote or where to do so.

That does not mean the risk is small. As Election Day approaches, Anthropic expects political and election-related questions to increase.

Election information within the conversation

Claude will continue showing a link to TurboVote, a nonpartisan resource from Democracy Works, when someone asks about voter registration, polling places or election dates.

In addition, starting in October, it will begin incorporating official voting information collected by Democracy Works from state and local sources into its responses. Someone will be able to ask which document they need to vote or where their polling place is and receive an answer specific to their state without leaving Claude.ai.

Anthropic also announced updated results during election night and the following days. When a user asks about the count, Claude will show the latest official figures available.

The difference matters: instead of responding with general or outdated information, the assistant will be able to direct each answer to specific data for the relevant state and point in time. Even so, it is worth checking the information with local election authorities if a decision depends on it.

More tests to detect political bias

Anthropic says its latest models, Claude Opus 5.5, Claude Sonnet 5.5 and Claude Haiku 5.5, scored 99% in its internal political impartiality evaluation. The test attempts to measure whether responses treat different political positions fairly.

The company also commissioned evaluations from three independent organizations: Future of Free Speech at Vanderbilt University; Foundation for American Innovation; and Collective Intelligence Project. Overall, their tests found that Claude responded substantively to requests about controversial topics and rarely refused to answer.

But they also identified specific problems. Claude sometimes added warnings or counterarguments that no one had asked for, including when drafting a message supporting a particular position. Another evaluation found that Claude Sonnet 4.5 softened arguments associated with right-wing positions more often than those linked to left-wing positions.

Anthropic says it is adjusting responses to direct questions, such as requests for a yes-or-no answer, and to requests for short texts. It is also working to better decide when nuance is necessary and when it interferes with what the user asked for.

The problem with campaigns carried out in stages

Anthropic's usage rules prohibit using Claude to deceive voters, spread false information, impersonate election officials or make it harder for people to vote. They also prohibit creating fake profiles, manufacturing mass comments to inflate a campaign's popularity or producing content for pages that appear to be independent news outlets.

The challenge does not always appear in an obvious request. Anthropic worked with the Institute for Strategic Dialogue and the McCain Institute to test its defenses in the United States, Israel and Brazil. The researchers broke possible manipulation operations into seemingly harmless tasks, such as researching audiences, creating a news page, drafting text and planning its distribution.

The system usually rejected a direct harmful request, but it could fail to detect an operation when it was spread across many requests that appeared legitimate on their own. Anthropic has added classifiers, automated systems that look for risk patterns, to better identify this type of coordinated abuse.

It also strengthened the detection of fake accounts, fraudulent news sites, election-related scams and operations linked to foreign state actors. Confirmed offending accounts may be blocked.

Cases Anthropic has already investigated

The company says it detected foreign operators of a fake news site who were using Claude to produce coverage of the U.S. legislative elections that favored narratives from Russia, Iran and North Korea.

It also found actors trying to exploit the elections to make money through fabricated news and other content designed to attract U.S. audiences and generate advertising revenue. Anthropic says it blocked the accounts involved and improved its systems to identify similar operations.

For you, the most visible change will be the ability to check official election data directly in Claude. The less visible, but more important, change will happen behind the scenes: systems that try to recognize when several ordinary tasks are part of a coordinated deception campaign. The real test will come as the volume of queries increases and political actors try to get around those defenses.