On October 14, 2026, Anthropic published an update describing steps it has taken to strengthen Claude’s safeguards for election‑related use ahead of the 2026 US midterm elections. The company says its aim is to keep Claude a useful tool for political conversations while preventing election misuse and misinformation.
How people use Claude for politics and elections
Anthropic used Anthropic Insights to analyze a random sample of 250,000 consumer conversations on Claude.ai from September 21–27, 2026, using privacy‑preserving methods. The company reports that roughly 5.75% of US conversations touch on politics, public policy or social issues, while about 0.51% are specifically about elections or voting.
Requests that explicitly ask Claude whom or what to vote for are rare in the sample (around 0.01% of US conversations), and queries that violate Anthropic’s election integrity policy appear even less often (about 0.002%). Most requests are informational: following news, researching candidates or policies, drafting documents, assisting with coursework, or weighing arguments for and against specific positions.
Connecting users to trusted sources
Ahead of Election Day, Anthropic added two methods to surface reliable voting information inside Claude conversations. The company continues to display a banner linking to TurboVote, a nonpartisan resource from Democracy Works, when users ask about voter registration, polling locations, or election dates.
Later in October, Claude began drawing official voting information collected by Democracy Works from state and local authorities into responses for US users, so state‑specific answers about where to vote or what ID to bring can appear directly in a conversation. Claude will also show up‑to‑date official election night results and subsequent counts when users ask for current vote totals.
Measuring and preventing political bias
Anthropic tests its models extensively for political bias before each model launch. The company reports that its latest models — Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 — scored 99% on its political even‑handedness evaluation. These models also engage more readily in political discussions and decline fewer requests, enabling users to get more of the information they seek.
Anthropic commissioned three independent reviewers to evaluate how Claude handles free expression and political content:
- The Future of Free Speech at Vanderbilt University tested Claude on hundreds of contentious requests.
- The Foundation for American Innovation compared responses to matched prompts from left and right across mainstream and extreme scenarios.
- The Collective Intelligence Project (CIP) surveyed more than 2,000 people in the US, UK and India to develop principles, select test topics, and create evaluation criteria, then tested agreement across demographic and cultural groups.
These evaluations generally found strong, substantive responses but also revealed weaknesses. The Future of Free Speech and CIP noted Claude rarely refused requests but sometimes inserted unsolicited caveats or counterarguments (for example, adding opposing views into an advocacy email). The Foundation for American Innovation found that Sonnet 4.5 did not always hedge evenly and could soften arguments for right‑leaning positions.
Anthropic says it has taken several steps in response: improving how Claude answers direct yes/no and short‑form questions, refining when and how the model adds caveats or counterarguments to ensure more consistent treatment across the political spectrum, and expanding bias evaluations to better reflect real conversations and to include questions about candidates and voting as well as policy.
Defining misuse and updating rules
Anthropic’s Usage Policy prohibits using Claude to deceive voters or disrupt elections, including spreading false information about candidates or voting, impersonating election officials, and suppressing turnout. The policy is updated annually; Anthropic says it announced the latest update the prior week and added examples to clarify the scope of its election rules.
Because many campaigns use Claude for legitimate work (drafting press statements and fundraising emails, debate prep, analyzing turnout data, building outreach materials), Anthropic published a guide explaining permissible uses and prohibitions — such as creating fake social profiles, mass‑producing comments to falsely inflate engagement, or writing for sites that misrepresent themselves as independent news outlets.
Testing and strengthening defenses
Anthropic deploys automated classifiers trained to detect election‑related misuse. To evaluate these safeguards, the company partnered with Institute for Strategic Dialogue (ISD) and the McCain Institute to pressure‑test protections in the US, Israel and Brazil. Analysts tried direct requests and also decomposed full disinformation campaigns into many small, seemingly benign tasks (audience research, building news sites, drafting and planning distribution). They found Claude reliably refused single harmful requests but could miss malicious campaigns split across many innocuous‑looking requests.
Anthropic has added new classifiers to better detect novel and multi‑step misuse — including attempts to build fake social accounts and news sites targeting political discussions or to run financially motivated election schemes. It also improved detection of election‑related misuse by foreign state actors, strengthened cyber protections for election infrastructure and campaigns, and built targeted safeguards for local risks identified in the US, Israel and Brazil.
Flagged activity is reviewed by threat‑intelligence and enforcement teams, and confirmed violations can result in account bans. Anthropic also routinely shares examples of unintended model behaviors observed during evaluations and internal use.
Findings and enforcement examples
Anthropic’s threat intelligence report highlights that political, state‑aligned and commercial actors are increasingly incorporating AI across influence operations, from planning and content creation to building distribution platforms. Because early stages of such operations sometimes appear on Claude, Anthropic can often intervene before material reaches the public.
The company described cases where foreign‑based operators used Claude to produce US midterm coverage favoring Russian, Iranian and North Korean narratives, and instances where actors used Claude to create fabricated news and other content aimed at American audiences to generate advertising revenue. Anthropic says it banned the accounts involved and improved detection for this class of abuse.
Looking ahead
Anthropic emphasizes its responsibility to ensure Claude provides accurate, balanced answers and is not used to undermine democratic processes. The company plans to continue investing in making Claude more even‑handed, in preventing and responding to misuse, and in transparency about its actions. In the coming months Anthropic expects to publish a more comprehensive report on how people used Claude throughout the election cycle to further strengthen its election evaluations and safeguards.
Notes: The referenced analysis used a random sample of 250,000 US Claude.ai consumer conversations from September 21–27, 2026, analyzed with privacy‑preserving methods; a separate review of 250,000 conversations generated the figure on election‑integrity policy violations.
