Anthropic has published its election playbook for the 2026 US midterms, and the most interesting number in it is a small one. Of the conversations it sampled, about 0.01% involved someone asking Claude whom or what to vote for. People, it seems, want a research assistant and not a political advisor, which is either reassuring or a quiet commentary on how much they trust software with their ballot.
The rest of the post covers what Anthropic changed since its earlier election update this year: new ways to get official voting information, a round of outside audits of political bias, and new defenses against a kind of abuse that the company's own testers found it was bad at spotting. It also reports a couple of real cases it has already shut down.
What people actually ask Claude about politics
Anthropic built its analysis with Anthropic Insights, looking at 250,000 conversations from a recent one-week period. The scope is US conversations on Claude.ai, the consumer app.
The headline figures:
| Category | Share of US Claude.ai conversations |
|---|---|
| Politics, public policy or social issues, broadly | about 5.75% |
| Specifically about elections or voting | about 0.51% |
| Asking Claude whom or what to vote for | about 0.01% |
| Violating election integrity policies | about 0.002% |
The broad political bucket covers following the news, academic or professional policy research, learning about candidates, and finding government services. Anthropic expects the political and election share to grow as Election Day approaches.
What people tend to ask is the sort of question that helps them reach their own view, such as where a candidate stands on a given issue. Requests that broke the company's election integrity rules were rarer still. Anthropic gives no breakdown beyond these percentages, so the sample tells us about proportions and not about the content of the bad requests.
Sending people to real answers
Chatbots are famously confident about things like polling hours, which is exactly where confidence is a liability. Anthropic's fix is to point at an actual source.
Since April, when someone asks Claude.ai about voter registration, polling places or election dates, a banner links to TurboVote, a nonpartisan resource from Democracy Works. That stays.
Two additions are new:
- Official voting information in the answer. Later this month, Claude on Claude.ai will draw on Democracy Works data for US users. Ask where to vote or what ID to bring, and you get an answer specific to your state without leaving the chat. The underlying information is publicly available material that Democracy Works collects from state and local election sources.
- Live results. On election night and the days after, Claude will show the latest official numbers directly in the conversation when someone asks for vote counts or results.
The design choice is worth noticing. Rather than trusting the model to recall election logistics, Anthropic is routing the question to a maintained source. That is a modest admission that a language model's memory is a poor substitute for a state election website.
Measuring bias, then asking others to check
Anthropic says it tests every model for political bias before launch. Its newest models, Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5, scored 99% on the company's own political evenhandedness evaluation, which it calls its best results to date. It also says these models engage more readily with political discussion and decline fewer requests.
A company grading its own homework and awarding itself 99% invites a raised eyebrow, and Anthropic seems to know it. It brought in three outside groups, each with a different method:
- The Future of Free Speech, an independent think tank at Vanderbilt University, tested Claude on hundreds of requests about contentious topics.
- The Foundation for American Innovation compared Claude's responses to matched prompts from the left and right, from mainstream policy debates to more extreme scenarios.
- The Collective Intelligence Project (CIP) surveyed more than 2,000 people in the US, UK and India to help build an evaluation. Participants proposed principles for handling sensitive topics (health through politics), picked topics to test, and helped define how responses should be judged. CIP then synthesized the input and checked for agreement across demographic and cultural groups.
On the first two evaluations, most responses were rated strong, meaning they engaged substantively with the request.
Where Claude still falls short
The more useful part of the post is the list of failures. Anthropic reports that the audits found real problems.
Both The Future of Free Speech and CIP found that Claude rarely refused requests, but sometimes tacked on unsolicited caveats and counterarguments. The example given is inserting an opposing view into an advocacy email, which is a bit like asking a speechwriter for a rally speech and getting a rebuttal stapled to the back.
The Foundation for American Innovation found that Claude Sonnet 4.5 did not always hedge evenly, and was more likely to soften arguments for right-leaning positions. Note the model name: that finding concerns Sonnet 4.5, while the 99% score belongs to the 5.5 generation. The post does not say whether the newer models were re-tested on the same matched-prompt comparison, so the two numbers should not be read as one continuous story.
Anthropic lists these responses:
- Direct questions. It improved how Claude handles yes-or-no questions and requests for shorter answers, where it found answers could be less balanced.
- Caveats. It is using the findings to refine when and how Claude adds caveats or counterarguments, aiming for more consistent treatment across the political spectrum.
- Uneven hedging. It is continuing to improve tests for it, building on evaluations that use matched left and right prompts.
- Wider coverage. It plans to expand its bias evaluations to look more like real conversations and to cover candidates and voting questions as well as policy issues.
That last item is an implicit admission that current evaluations lean toward policy debate, while real users are asking about candidates and ballots.
Where the line on misuse sits
Anthropic's Usage Policy bars using Claude to deceive voters or disrupt elections. That covers spreading false information about candidates or how to vote, impersonating election officials, and trying to suppress turnout. The policy is updated each year to track model capabilities and customer feedback. The latest update, announced last week, clarifies what the election rules cover and what they do not, with new examples of where the lines fall.
Campaigns are not the enemy here, and the company says so. It recently published a guide for political campaigns covering legitimate uses such as:
- drafting press statements and fundraising emails
- prepping for debates
- analyzing turnout data
- building voter outreach materials
The guide also spells out what is off limits: fake social media profiles posing as real people, mass-produced comments to inflate a campaign's engagement, and content for news sites or pages that falsely claim independence.
The attack that looks like a to-do list
Anthropic uses automated classifiers trained to detect election-related abuse. To see whether they hold up, it hired the Institute for Strategic Dialogue (ISD) and the McCain Institute, both independent research organizations, to pressure-test protections in the US, Israel and Brazil.
The analysts tested whether Claude would help with deceptive campaigning, misinformation and other abuse. They started with direct requests, then split whole campaigns into smaller tasks that look innocent alone: audience research, building news sites, drafting content, planning distribution.
The result is the sort of finding that sounds obvious once someone says it. Claude reliably refused a single harmful request, but could miss a malicious campaign spread across many harmless-looking ones. Nobody running an influence operation needs to type "help me run an influence operation." A project plan will do.
Anthropic's response includes:
- new classifiers aimed at novel and multi-step misuse, including attempts to build fake social media accounts and news sites for political discussions, and financially motivated election schemes
- better detection of election-related misuse by foreign state actors
- stronger cyber protections for election infrastructure and campaigns
- targeted safeguards for localized risks found in the US, Israel and Brazil, all of which have major national elections this year
Threat intelligence and enforcement teams review flagged activity and act when a violation is confirmed, including banning accounts. Anthropic adds that it routinely shares examples of unintended model actions or behavior seen in its evaluations and internal use of Claude.
What Anthropic has already caught
The company says its recent threat intelligence report reflects a wider shift: political, state-aligned and commercial actors now use AI in nearly every part of an influence operation, from planning and content creation to building the platforms that spread a message. Because those early stages happen on Claude, Anthropic says it often has a chance to intervene before anything reaches the public.
Two cases are described:
- A fake news site. Foreign-based operators used Claude to produce US midterm coverage favoring Russian, Iranian and North Korean narratives.
- Ad-revenue farming. Threat actors tried to exploit the election for money, using Claude to produce fabricated news and other content built to attract American audiences and generate advertising revenue.
Anthropic banned the accounts in both cases and says it improved its ability to detect this type of abuse. The post gives no counts of accounts or volume of content, so the scale is unknown.
What comes next
Anthropic says more people are turning to Claude for election information than ever before, and that it will keep working on evenhandedness, misuse prevention and transparency. It says the work extends to elections beyond the US. In the coming months it hopes to publish a fuller look at how people used Claude across the whole election cycle, to sharpen its evaluations and safeguards.
The interesting promise is that last one. The company has just told us its first-pass bias tests were too narrow and its first-pass abuse detection missed things that looked boring. A post-election review might be where we learn whether the fixes worked, or just which new boring-looking tasks slipped through.
Sources
- Preparing for the 2026 US midterms — Anthropic