Anthropic published a round-up of its election safeguards on 14 October 2026, ahead of this year's US midterm elections, with one figure summing up the picture: 5.75% of Claude.ai conversations in the United States touch on politics, public policy or social issues, and only 0.51% are specifically about elections or voting. Requests that break its election integrity rules show up in roughly 0.002% of conversations.
What was announced
The measurement comes from a random sample of 250,000 US consumer conversations on Claude.ai, collected between 21 and 27 September 2026 and classified, the company says, using a privacy-preserving method. Beyond the 5.75% and 0.51% figures, Anthropic notes that asking the model who or what to vote for is vanishingly rare: around 0.01% of conversations, and it makes clear that this is the top end of a 95% confidence interval, because the case is too infrequent to pin down precisely. The 0.002% violation figure comes from a separate analysis, also of 250,000 conversations, reviewed against its election integrity policy.
What it does see often, it says, is informational use: following the news, researching where candidates stand, drafting documents, coursework, and weighing arguments for and against a policy. It also warns that the share of political questions will rise as polling day approaches.
On top of that come two product changes. First, this month Claude will start pulling official voting information from Democracy Works into its answers for US users. Anyone asking where they vote or which ID to bring to the polling place will get a state-specific answer without leaving the chat, built from public data the organisation gathers from state and local election authorities. The banner introduced in April, pointing to the same group's non-partisan TurboVote service, stays in place. Second, on election night and the days after it, Claude will show up-to-date official counts and results inside the conversation.
On political bias, Anthropic says Opus 5.5, Sonnet 5.5 and Haiku 5.5 are its most balanced models so far, scoring 99% on its in-house evenhandedness evaluation, and that they decline fewer political requests than earlier versions. To stress-test that, it commissioned three external reviews with different methods: a free-expression think tank at Vanderbilt University ran hundreds of requests on contentious subjects; the Foundation for American Innovation compared answers to matched questions framed from the left and from the right; and the Collective Intelligence Project surveyed more than 2,000 people across the United States, the UK and India to build, over three stages, the principles and the scoring criteria for model answers.
All three found problems. Two saw that the model almost never refuses, but slips in caveats and counterarguments nobody asked for — dropping the opposing view into an advocacy email, for instance. The third found that Sonnet 4.5 did not spread its hedging evenly and tended to soften right-leaning arguments more. Anthropic says it has reworked how the model handles yes-or-no questions and requests for short answers, the two formats where the imbalance showed up most.
Where it comes from
The post links back to the safeguards update the company published earlier this year and to the Usage Policy rewrite announced the week before, which takes effect on 12 November and has already been covered on this blog. That rewrite spells out what the election rules do and do not cover, and adds examples. In parallel, Anthropic has issued a guide for political campaigns: it accepts that they use Claude daily for legitimate work — press statements, fundraising emails, debate prep, turnout analysis — and draws the line around the banned uses, from fake profiles posing as real people to mass-produced comments inflating engagement and writing for sites that falsely present themselves as independent outlets.
The defensive side was tested with two independent organisations, the Institute for Strategic Dialogue and the McCain Institute, in the United States, Israel and Brazil, all three holding national elections this year. The result is the most interesting data point in the announcement: Claude blocks a direct harmful request reliably, but may miss the whole operation when it is sliced into tasks that look harmless on their own — audience research, standing up a news site, drafting copy, planning distribution. Hence the new classifiers aimed at multi-step misuse.
As for real cases, the company cites foreign operators of a fake news site that used Claude to generate midterm coverage aligned with Russian, Iranian and North Korean narratives, and financially motivated actors producing content to attract American audiences and monetise it through advertising. Those accounts were banned.
What it means
For anyone building products on top of models, the stress-test finding travels well beyond elections: a filter that judges each request in isolation will not catch abuse spread across ten perfectly acceptable requests. Useful defence looks at the session, the account and the pattern, not the single message.
The second reading is architectural. Anthropic is not trying to make the model remember where each county votes: it connects an authoritative source and injects it into the answer. That is the pattern worth applying to any volatile, verifiable data — opening hours, fares, regulations: retrieval from a source with a named owner, not model memory.
And the third concerns tone. The audits do not flag censorship, they flag excess hedging. An assistant that bolts counterarguments onto work nobody asked it to balance is both annoying and, in effect, slanted. Worth rereading your own system prompts with that lens.
What the announcement leaves open
The 99% evenhandedness score comes from Anthropic's own evaluation, not an independent exam, and the company admits it is rebuilding that evaluation to cover candidates and voting questions, not just policy debates. The usage percentages come from one specific week in September and from automated classification: they are not a picture of the whole cycle, and the company itself expects the numbers to climb.
Nor is there detail on exactly what the Democracy Works feed covers, how often it refreshes, or what happens if a state changes its rules at the last minute. No figures are given for the abuse cases: not how many accounts were banned, nor how much content reached the public before detection. And the fix for uneven hedging is described as work in progress — better handling of short answers, more consistent criteria for caveats — with no follow-up measurement showing that the imbalance found in Sonnet 4.5 has gone.
The full announcement is in Anthropic's note on the 2026 midterms. The company says it hopes to publish a broader look at how Claude was used across the election cycle in the coming months.