Anthropic Midterm Safeguards: Only 0.01% of Claude Chats Ask Who to Vote For

Anthropic has published its plan for the US midterms, and the most useful part is not the policy language but the numbers. In a post dated October 14, 2026, the company reports that about 5.75% of US conversations on Claude.ai touch politics, public policy or social issues, about 0.51% are specifically about elections or voting, and only about 0.01% ask Claude whom or what to vote for. Requests that broke its election-integrity policy were rarer still, at about 0.002%.

Those figures come from a privacy-preserving analysis of roughly 250,000 US consumer conversations sampled from September 21 to 27, 2026, built with the Anthropic Insights tooling. Alongside them, Anthropic describes new product behavior (official voting data inside answers, election-night results), fresh bias testing, and a case of a foreign-run fake news site it shut down. This post walks through what is confirmed, what is Anthropic's own measurement, and what builders and voters should take from it.

TL;DR: what changed for the midterms

How do people actually use Claude for politics?

The first thing the post does is replace speculation with a measurement. The fear in most election-and-AI discussions is that chatbots become persuasion machines: ask a question, get told how to vote. Anthropic's sample points the other way. The overwhelming majority of political traffic is informational: people follow the news, research candidates, draft documents, help with coursework and weigh arguments for and against a policy.

Each dot in Anthropic's chart represents one in every 10,000 conversations. Because vote-recommendation requests are so rare, the company shows the upper end of a 95% confidence interval for that group rather than a point estimate, which is the honest way to plot a tiny number. Anthropic also expects the political and election share to climb as Election Day approaches, so the September sample is a floor, not a forecast of peak load.

Two caveats matter for anyone quoting these figures. First, the sample is US consumer Claude.ai traffic only; it says nothing about API usage, enterprise deployments, or how developers embed Claude in their own apps. Second, classification was done by Claude itself, using a privacy-preserving clustering method, so the percentages inherit whatever errors that classifier has. Anthropic labels the examples in the chart as illustrative only.

The practical read: a campaign staffer drafting a fundraising email, a student comparing two ballot measures and a voter asking where their polling place is are the typical political users. The persuasion-chatbot scenario is, in this dataset, a rounding error.

What is Anthropic adding to Claude for election season?

Two product changes are the concrete news here.

Official voting information inside answers

Claude.ai already shows a banner linking to TurboVote, a nonpartisan resource from Democracy Works, when people ask about voter registration, polling places or election dates. Anthropic calls that the industry standard. What is new, starting later this month for US users on Claude.ai, is that Claude will pull official voting information from Democracy Works straight into its responses. Someone asking where to vote or which ID to bring should get an answer specific to their state without leaving the conversation. Anthropic says the data draws on public information Democracy Works collects from state and local election sources.

This is the right design choice for a boring but important reason: voting logistics are exactly the kind of fact a model can get subtly wrong, and wrong logistics (a wrong deadline, a wrong ID rule) are a form of voter suppression even when nobody intends it. Routing those answers to a maintained, sourced dataset is safer than relying on whatever the model memorized in training. The same pattern, grounding a high-stakes answer in an authoritative feed, is worth copying in any app you build where a wrong answer has consequences. Our guide to how Claude's usage policy changes land on November 12 covers the policy side of the same push.

Live results on election night

Claude will also begin showing up-to-date results on election night and in the days that follow. When people ask for current vote counts or results, they will be shown the latest official numbers directly in the conversation rather than a model's guess. Anthropic does not name the results provider in the post, so treat that detail as unconfirmed until the feature ships. What the post does commit to is the behavior: official numbers, surfaced in chat.

How even-handed is Claude, and where does it still slip?

Anthropic says it tests its models for political bias before every launch, and that Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 are its most even-handed models to date, each scoring 99% on its political even-handedness evaluation. It adds that the latest models engage more readily with political discussions and decline fewer requests. If you follow the model line, those are the same generation we covered at launch in Claude Opus 5.5's benchmarks and pricing and Haiku 5.5's pricing and cache cut.

The more interesting part is the outside review. Anthropic asked three independent groups to examine how Claude handles free expression, each with a different method:

• The Future of Free Speech (a Vanderbilt University think tank) tested Claude on hundreds of requests on contentious topics.
• The Foundation for American Innovation compared responses to matched prompts from the left and right, from mainstream policy debates to extreme scenarios.
• The Collective Intelligence Project surveyed more than 2,000 people in the US, UK and India, who proposed principles for sensitive topics, picked topics to test and helped define the criteria for judging answers.

Most responses in the first two evaluations were rated strong and substantively engaged. But Anthropic is candid about the misses. The Future of Free Speech and the Collective Intelligence Project both found Claude rarely refused, yet sometimes added unsolicited caveats and counterarguments, such as inserting an opposing view into an advocacy email. 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 that Sonnet 4.5 is an older model than the 5.5 family Anthropic is promoting for the 99% score, so the two findings are not a like-for-like comparison.

Anthropic's stated fixes: better handling of direct yes-or-no questions and requests for shorter answers (where balance suffered), better judgment about when to add caveats or counterarguments, and expanded matched-prompt testing that covers candidates and voting questions, not only policy issues.

For builders this is the practical lesson. If you ship a product that writes persuasive content, the failure mode is not refusal, it is uneven hedging: one side gets a lecture, the other gets a draft. You can test for it yourself with matched prompt pairs, the same technique these reviewers used. For a broader look at where models sit politically, see our piece on AI models and the political compass.

What counts as election misuse under Claude's rules?

Anthropic's usage policy prohibits using Claude to deceive voters or disrupt elections: false claims about candidates or how to vote, impersonating election officials, suppressing turnout. The company says this year's policy update, announced last week, clarifies what the election rules cover and adds examples. Our write-up of that update is in Anthropic's 2026 usage policy.

Campaigns get a guide too. Anthropic says many campaigns use Claude for legitimate work such as press statements, fundraising emails, debate prep, turnout-data analysis and voter-outreach materials. The prohibited list includes:

1. Creating fake social media profiles that pose as real people.
2. Mass-producing comments to inflate a campaign's engagement.
3. Writing content for news sites and pages that falsely claim to be independent.

That third item connects directly to the case Anthropic describes below, and it is the line that will matter most in practice, because the same drafting skill that writes a legitimate op-ed writes a fake local news story.

What did Anthropic find in the wild?

Anthropic says its most recent threat intelligence report reflects a broader 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 the early stages of an operation often happen on Claude, the company says it often has a chance to stop efforts before they reach the public.

Two midterm-specific cases are described, both without naming accounts or sites:

• Foreign-based operators of a fake news site used Claude to produce US midterm coverage favoring Russian, Iranian and North Korean narratives.
• Other actors used Claude to fabricate news and content aimed at American audiences, purely to generate advertising revenue.

Anthropic banned the accounts involved and says it improved its ability to detect this kind of abuse. These are Anthropic's own disclosures; the company gives no site names, volumes or reach, so there is nothing independent to verify yet. For comparison, OpenAI described similar takedowns of Russia- and Iran-linked fronts in its own influence-operation disruption report. The pattern across labs is consistent: the content is cheap to produce, so the contest moves to detection and to the distribution platforms.

How were the defenses tested?

Anthropic deploys automated classifiers trained to flag election-related abuse. To test them, it partnered with the Institute for Strategic Dialogue (ISD) and the McCain Institute to pressure-test protections in the US, Israel and Brazil, three countries with major national elections this year. Analysts started with direct requests, then split full campaigns into smaller tasks that look harmless alone: audience research, building news sites, drafting content, planning distribution. That decomposition test is the realistic one, since a determined operator will never ask for "a disinformation campaign" in one prompt.

The result is the most candid line in the post: Claude reliably refused a single harmful request, but could miss a malicious campaign when it was spread across many individual requests that looked harmless on their own. Anthropic says it then strengthened detection of coordinated misuse, including attempts to build fake social accounts and news sites and financially motivated election schemes, and improved its ability to spot foreign state-actor activity. It also says it hardened cyber protections for election infrastructure and campaigns and built targeted safeguards for localized risks in the US, Israel and Brazil.

What this means if you build with Claude

If you build on the API, the consumer-app protections described here (the Democracy Works integration, election-night results, the TurboVote banner) are Claude.ai features. Anthropic's post does not say they apply to API calls, so do not assume your app inherits them. If your product answers voting-logistics questions, link to an official source such as your state election office or Democracy Works yourself, and keep a human-reviewed fallback.

Three habits follow from the post:

1. Ground factual election answers. Deadlines, ID rules and results should come from a maintained source, not model memory.
2. Test for uneven hedging. Run matched left and right prompts through your pipeline and compare caveat rate, length and tone, not only refusals.
3. Watch the decomposition risk. If your agent can research audiences, build sites and draft copy, a bad actor can chain those steps. Rate limits and review on bulk publishing matter more than any single-prompt filter.

For teams running coding or content agents, the broader point about stopping dangerous actions before they happen is what AgentBeam, the agent security platform from the explainx.ai team, is built for.

What to watch next

Anthropic says it hopes to share, in the coming months, a more comprehensive look at how people used Claude throughout the election cycle, with the goal of strengthening its election evaluations and safeguards. That follow-up is the one to read: this post gives a one-week snapshot, and the interesting data is the curve as Election Day nears.

The post also lands in a week of heavy AI policy news. Anthropic's Presidential Engagement Program for 2028 candidates shows the company is engaging politicians directly, while regulators are circling the labs in the FTC probe of OpenAI and Anthropic. Taken together, election season is becoming the first large public test of whether lab safeguards work outside the lab.

Our read: the 0.01% number is the headline, but the product change is the substance. Moving voting logistics out of the model's memory and into an official feed is the kind of unglamorous engineering that actually prevents harm. The open questions are whether the same grounding reaches API users, who supplies the results feed, and whether the red-team findings Anthropic summarizes hold up when outside researchers can see the details.

Related reading

• Anthropic Insights: independent research on Claude usage
• Anthropic's 2026 usage policy, effective November 12
• Anthropic threat intelligence report, September 2026
• Presidential Engagement Program for 2028 candidates
• OpenAI disrupts Russia and Iran influence operations
• AI models and the political compass
• FTC probe of OpenAI and Anthropic
• Official: Democracy Works, Anthropic's midterms post

Figures and policy details are accurate as of October 14, 2026 and come from Anthropic's own post; features described as "later this month" had not shipped at publication.