Anthropic analyzed 250 thousand Claude chats ahead of the US elections

Anthropic has published a report on how it is preparing Claude for the 2026 US midterm elections. The report rests on an analysis of about 250 thousand conversations from American Claude.ai users during the week of September 21 to 27. Conversations about politics, public policy and social issues make up 5.75% of all traffic, while 0.51% are about elections and voting. The company expects both shares to grow by Election Day.

What people actually ask

The data was collected with Anthropic's Insights tool. It clusters conversations from the bottom up and preserves privacy: Claude itself does the classification. In the report's chart, each dot corresponds to one conversation out of 10 thousand.

The most notable finding concerns requests for advice on who or what to vote for. Such requests make up about 0.01%. Because the sample is small, Anthropic shows the upper bound of the 95% confidence interval. More often, people use Claude to work things out for themselves: to learn a candidate's position on a specific issue, understand a geopolitical event, prepare a document, do homework or compare arguments for and against a policy. Requests that violate the company's election integrity policy are rarer still, at about 0.002% of dialogues.

Official data right in the chat

Since April, when users ask about voter registration, polling places and dates, Claude.ai has shown a banner linking to TurboVote, a nonpartisan resource run by Democracy Works. Anthropic calls this practice an industry standard. Two new mechanisms are now being added.

• Later in October, Claude.ai will begin pulling official voting information from Democracy Works into its answers for users in the US. A question such as "where do I vote" or "what ID do I bring to the polling place" will get a state-specific answer without sending the user to a third-party site. The sources are public data from state and local election authorities, compiled by Democracy Works.
• On election night and in the following days, Claude will show current official results and vote counts directly in the conversation.

Political even-handedness and external audit

Anthropic tests its models for political bias before every release. According to the company, Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 are the most even-handed models in its history, scoring 99% on its own political even-handedness evaluation. The new models engage in political discussions more readily and refuse requests less often.

To avoid relying solely on internal metrics, Anthropic brought in three independent groups, each with its own methodology. Future of Free Speech, an independent think tank at Vanderbilt University, ran Claude through hundreds of prompts on contested topics. The Foundation for American Innovation compared responses to paired "left" and "right" prompts, ranging from mainstream political discussions to extreme scenarios. The description of the third group and the final results are cut off in the available portion of the report, so we stick to what is known for certain.

Key figures

• Sample: about 250 thousand dialogues from American Claude.ai users, September 21–27, 2026
• Politics, public policy, social topics: 5.75%
• Elections and voting: 0.51%
• Requests for advice on who to vote for: about 0.01%
• Violations of the election integrity policy: about 0.002%
• Opus 5.5, Sonnet 5.5 and Haiku 5.5 score on political even-handedness: 99%

Practical takeaway

For American voters, the main change is that Claude will act as a reference for voting procedure. It will answer with state-specific details rather than generalities, drawing on Democracy Works data. For critical details, such as ID requirements, it makes sense to check the local election office's website: Claude only retells compiled public sources. For developers and political teams, the report shows what to expect from the model. It will help break down positions and prepare documents, but it will not act as a campaigner, and requests for manipulation fall under separate restrictions. These rules will be tested in real conditions on Election Day.