Anthropic Studied a Quarter Million Claude Chats Before the Midterms

With the midterm elections less than two weeks away, Anthropic has opened its notebook on Claude Chats. The AI company published an October 14 breakdown of how Americans actually use its chatbot around politics, drawn from a random sample of 250,000 United States consumer conversations collected between September 21 and September 27, 2026. The analysis, run through Anthropic Insights in a way the company describes as privacy-preserving, is paired with a set of new safeguards meant to keep election answers accurate when voters start asking the questions that matter most.

The headline finding is how little of it is about picking a side. Conversations touching broadly on politics, public policy or social issues made up about 5.75% of United States chats on Claude.ai during the sample week, according to the company's figures. Only around 0.51% were specifically about elections or voting. Requests asking the chatbot whom or what to vote for appeared in roughly 0.01% of conversations — a figure so small that Anthropic reported the upper end of a 95% confidence interval rather than a point estimate. Requests that violated its election integrity policies were rarer still, showing up in about 0.002% of chats in a separate review of the same size.

Most political traffic, Anthropic says, is informational. People use the assistant to follow geopolitical events, research candidates, draft documents, work through coursework, navigate government services and weigh arguments for and against policies. The company expects the election share to grow as Election Day approaches, and it is betting that the people who do show up with voting questions will get answers anchored to official sources rather than model memory.

Official voting data moves inside the chat

Since April 2026, Claude.ai has shown users a banner linking to TurboVote, the nonpartisan voter tool run by Democracy Works, whenever they ask about registration, polling places or election dates. Later this month, that changes in a meaningful way: Claude will begin drawing official voting information from Democracy Works directly into its responses for United States users. Questions like where to vote or what identification to bring would receive state-specific answers inside the conversation itself, built from public data the nonprofit collects from state and local election sources. Claude will also show up-to-date official vote counts and results on election night and in the days after.

The shift matters because this is a domain where a wrong answer has immediate consequences. A wrong deadline or a wrong identification rule can cost someone their vote. Anthropic acknowledged back in 2024 that Claude was not being retrained often enough to answer real-time election questions reliably — a gap the new setup is designed to close by grounding the highest-stakes answers in data maintained by election administrators rather than the model's own training cutoff.

Three outside reviews, and what they found

Anthropic says it tests its models for political bias before every launch and reports that its newest generation scored 99% on its internal political even-handedness evaluation, which it describes as its most even-handed result to date. Three independent groups then reviewed how Claude handles free expression, each with a different method. The Future of Free Speech, a think tank at Vanderbilt University, ran hundreds of requests on contentious topics. The Foundation for American Innovation compared responses to matched prompts from the political left and right. The Collective Intelligence Project surveyed more than 2,000 people across the United States, the United Kingdom and India in a three-stage process to set shared principles for judging responses.

The reviews were mostly strong but not clean. According to reporting by Unite.AI, the first and third groups found Claude sometimes added unsolicited caveats and counterarguments, such as inserting an opposing view into an advocacy email, while the Foundation found one model did not hedge evenly, softening arguments for right-leaning positions more often. Anthropic says it has improved answers to direct yes-or-no questions and shorter requests, is refining when the assistant adds caveats, and is expanding its bias evaluations to cover candidates and voting questions as well as policy issues.

Red-teaming the campaigns that hide in plain sight

To pressure-test its defenses, Anthropic partnered with the Institute for Strategic Dialogue and the McCain Institute to probe its protections in the United States, Israel and Brazil — three countries holding major national elections in 2026. Analysts started with direct harmful requests, then broke full influence campaigns into smaller tasks that could look harmless alone: audience research, building news sites, drafting content. The testers found Claude reliably refused a single harmful request but could miss a malicious campaign fragmented across many individually benign-looking requests, a blind spot the company says it is now addressing with classifiers built to catch novel and multi-step misuse.

Enforcement is already producing cases. Anthropic says foreign-based operators of a fake news site used Claude to produce United States midterm coverage favoring Russian, Iranian and North Korean narratives, and that other actors used the assistant to fabricate news designed to attract American audiences and generate advertising revenue. The company says it banned the accounts involved. Its September threat intelligence report documented nine disrupted influence operations and assessed that most of the resulting content drew little authentic engagement, in part because Anthropic often disrupts operations during the production stage, before the material reaches the public.

The company also tightened its written rules. A revised usage policy published earlier in October clarifies what the election rules cover and what they do not, prohibiting uses such as deceiving voters, spreading false information about candidates or voting procedures, impersonating election officials and suppressing turnout. The updated policy takes effect on November 12. A separate October guide for political campaigns lists permitted everyday uses — drafting press statements and fundraising emails, preparing for debates, analyzing turnout and precinct data — and prohibited ones, including fake personas posing as real people, mass-produced comments inflating engagement, false posts about when or where to vote and undisclosed synthetic media impersonating a candidate.

Why this matters to a first-time voter

For young voters, the practical takeaway is simple: Claude Chats are already where many do homework and research, and that same window is about to become a front door to official voting logistics. Typing "where do I vote" should soon return the actual polling place and identification rules for your state, drawn from the same data election administrators publish, rather than a generated guess. On election night, the results shown will be the official counts, not a model's reconstruction of them. That will not settle arguments about bias in AI — the outside reviews show why that debate will continue — but it does narrow the most dangerous failure mode, which is a confident wrong answer about how to cast a ballot.

Anthropic says it hopes to publish a fuller account of how people used Claude across the election cycle in the coming months, and describes the midterm work as a template for elections around the world. Whether voters trust an AI assistant with the mechanics of democracy will depend less on this week's statistics than on whether the answers hold up under the pressure of an actual election night. For related coverage of how the same company is trying to shape the rules in Washington, see Anthropic Drafted the AI Safety Law It Wants Congress to Pass, and for how another chatbot is handling private conversations, see ChatGPT Locked Chats Keep Private Threads Behind a PIN. On the misinformation side of this election, read Maricopa County Ballot Video Sparks Fresh Election Conspiracy.

Sources: Unite.AI on Anthropic's October safeguards announcement; Pulse2 on the Claude election safeguards; MadRobot on the usage analysis and Democracy Works integration.