Anthropic Analyzed 250,000 Chats Ahead of the 2026 Elections

Ahead of the US elections in fall 2026, Anthropic published an updated report on how Claude is protected against political manipulation and disinformation. The company analyzed 250,000 Claude.ai user chats from September 21 to 27, 2026, to understand how people use AI for political purposes and what risks deserve attention.

How users engage with Claude on politics

The numbers point to predictable demand. Politics, political issues, or social problems come up in 5.75% of all American conversations on Claude.ai. Elections and voting specifically account for 0.51% of chats. Requests like "tell me who to vote for" or "which party should I choose" appear in just 0.01% of cases. Direct violations of election-integrity policy are even rarer, at 0.002%. The bulk of activity is informational: users look up candidates' positions on issues, research geopolitical events, work on term papers, and weigh arguments for and against a given policy.

Integrating official information sources

Anthropic added a direct link to TurboVote, a nonprofit service run by Democracy Works. When a user asks about voter registration, polling locations, or election dates, Claude will point them to TurboVote.

Starting in late September 2026, Claude embeds official election information directly into its Claude.ai responses. If a user asks where to vote or which document to bring to the polling place, the AI gives a precise, state-specific answer. Claude draws on data that Democracy Works collects from official state and municipal sources. During elections, Claude also displays real-time results.

Political balance and bias testing

Anthropic's newest models — Opus 5.5, Sonnet 5.5, and Haiku 5.5 — score 99% on the company's political-balance evaluation. They are more willing to discuss political topics and refuse less often than their predecessors.

Anthropic brought in three independent organizations to evaluate how Claude handles free-speech questions. Future of Free Speech, a research center at Vanderbilt University, tested Claude on hundreds of prompts about contentious topics. The Foundation for American Innovation compared Claude's responses to identical questions framed from left-leaning and right-leaning perspectives, ranging from mainstream debates to extreme scenarios. The Collective Intelligence Project surveyed more than 2,000 people across the US, UK, and India to help develop principles for handling sensitive subjects, from healthcare to politics.

The testing surfaced some issues: Claude rarely refuses outright, but it sometimes adds unnecessary caveats and counterarguments — for example, inserting an opposing viewpoint into the text of a lobbying letter. Sonnet 4.5 softened arguments inconsistently, treating right-leaning positions more gently. Anthropic is working to improve direct-question responses, make caveat insertion consistent across the political spectrum, and expand testing to cover candidates and voting-related questions.

Defenses against manipulation and new classifiers

Anthropic tested its safeguards against policy violations in partnership with the Institute for Strategic Dialogue, using case studies from the US, Israel, and Brazil. Researchers tried to get Claude to assist with manipulation, disinformation, and other forms of election sabotage. The conclusion: Claude refuses direct harmful requests, but it can miss a multi-step campaign in which each individual step — building an audience, setting up a news site, writing content, distributing it — looks harmless on its own.

Based on these findings, Anthropic added new classifiers to detect multi-step attacks, attempts to create fake social media accounts, politically slanted fake news sites, election-linked financial schemes, and foreign-government interference. The company also strengthened cyber defenses for election infrastructure and campaigns.

Anthropic's threat-intelligence teams detected and blocked several campaigns. Foreign operators running a fake news site used Claude to produce content about the 2026 US elections that favored Russian, Iranian, and North Korean narratives. Other threat actors used Claude to generate fabricated news stories aimed at attracting an American audience and earning ad revenue. All of the accounts involved have been banned, and defenses have been strengthened based on the new patterns identified.

Anthropic says it is committed to transparency and will continue investing in improving Claude's balance, preventing manipulation, and communicating openly with users. Over the course of the election cycle, the company plans to publish a comprehensive analysis of how Claude is used for political purposes.