Anthropic on October 14, 2026 detailed strengthened election safeguards forward of the US midterms, releasing a privacy-preserving evaluation of 250,000 Claude conversations alongside plans to deliver official voting info and reside election outcomes into Claude.ai.
How People Use Claude for Politics
The evaluation drew on a random pattern of 250,000 US Claude.ai shopper conversations from September 21 to 27, 2026, labeled in a privacy-preserving means via Anthropic Insights. Anthropic reported that conversations touching broadly on politics, public coverage, or social points accounted for about 5.75% of all US conversations on Claude.ai, its shopper app, and that about 0.51% involved elections or voting particularly. Requests asking Claude whom or what to vote for appeared in about 0.01% of conversations, the corporate mentioned, noting that as a result of such requests are uncommon it displayed the higher finish of a 95% confidence interval for that group. Requests violating its election integrity insurance policies appeared in about 0.002% of conversations, measured in a separate evaluation of 250,000 conversations. Anthropic mentioned it expects the share of political and election-related queries to extend as Election Day approaches, and described most noticed requests as informational: researching candidates, drafting paperwork, following geopolitical occasions, helping with coursework, and weighing arguments for and in opposition to insurance policies.
Official Voting Knowledge and Election-Night time Outcomes
Since April 2026, Claude.ai has displayed a banner linking to TurboVote, a nonpartisan useful resource from Democracy Works, when customers ask about voter registration, polling locations, or election dates, a characteristic first described within the firm’s April 2026 election-safeguards replace. Later in October 2026, Claude will start drawing official voting info from Democracy Works immediately into its Claude.ai responses for US customers, in order that questions reminiscent of the place to vote or what identification to deliver to a polling place obtain state-specific solutions with out leaving the dialog. The responses draw on publicly out there info that Democracy Works collects from state and native election sources. Claude may also start exhibiting up-to-date official vote counts and election outcomes on election evening and within the days that comply with.
Bias Testing and Three Exterior Evaluations
Anthropic mentioned it checks its fashions for political bias earlier than each launch and reported that Claude Opus 5.5, Sonnet 5.5, and Haiku 5.5 scored 99% on its political even-handedness analysis, describing them as its most even-handed fashions to this point.
Three exterior teams reviewed Claude’s dealing with of free expression, every with a special methodology. The Way forward for Free Speech, an unbiased suppose tank at Vanderbilt College, examined Claude on lots of of requests on contentious matters. The Basis for American Innovation in contrast how Claude responded to matched prompts from the political left and proper. The Collective Intelligence Challenge surveyed greater than 2,000 individuals within the US, UK, and India in a three-stage course of to ascertain shared ideas for judging responses. Anthropic reported that the majority responses have been rated sturdy on the primary two evaluations, however the critiques additionally documented shortcomings: The Way forward for Free Speech and the Collective Intelligence Challenge discovered Claude typically added unsolicited caveats and counterarguments, reminiscent of inserting an opposing view into an advocacy electronic mail, and the Basis for American Innovation discovered Claude Sonnet 4.5 didn’t hedge evenly, softening arguments for right-leaning positions extra usually. Anthropic mentioned it has improved responses to direct yes-or-no questions and shorter-answer requests, is refining when Claude provides caveats, and is increasing its bias evaluations to cowl candidates and voting questions in addition to coverage points.
Utilization Guidelines and a New Marketing campaign Information
The corporate’s Utilization Coverage prohibits utilizing Claude to deceive voters or disrupt elections, together with spreading false details about candidates or voting procedures, impersonating election officers, and suppressing turnout via deception or intimidation; the up to date coverage takes impact November 12, 2026. Anthropic revised the coverage earlier in October 2026 to make clear what the election guidelines cowl and so as to add examples.
A brand new October 2026 information to utilizing Claude in election campaigns lists permitted makes use of reminiscent of drafting press statements and fundraising emails, getting ready for debates, analyzing turnout and precinct information, triaging constituent correspondence, and creating accessible outreach supplies in large-print, screen-reader-friendly, and translated variations. Prohibited makes use of embrace pretend personas posing as actual individuals, mass-produced feedback inflating a marketing campaign’s engagement, information pages that falsely current themselves as unbiased, false posts about when or the place to vote, and undisclosed artificial media impersonating a candidate, with enforcement as much as and together with an account ban. The information notes that customers stay accountable for compliance with election legal guidelines, together with disclaimer necessities on public communications.
Crimson-Teaming, Classifiers, and Enforcement
To check its defenses, Anthropic partnered with the Institute for Strategic Dialogue and the McCain Institute, each unbiased analysis organizations, to pressure-test its protections within the US, Israel, and Brazil, which all maintain main nationwide elections in 2026. Anthropic mentioned analysts started with direct dangerous requests, then broke full campaigns into smaller duties that would look innocent on their very own, reminiscent of viewers analysis, constructing information websites, and drafting content material. The testers discovered that Claude reliably refused a single dangerous request however might miss a malicious marketing campaign fragmented throughout many individually benign-looking requests.
Constructing on these findings, Anthropic mentioned it added classifiers to determine novel and multi-step misuse, together with makes an attempt to construct pretend social media accounts and information websites geared toward political discussions or to run financially motivated election schemes. It additionally enhanced detection of election-related misuse by overseas state actors, strengthened cyber protections for election infrastructure and campaigns, and constructed focused safeguards for localized dangers within the three nations.
The corporate additionally reported enforcement outcomes: in a single case, foreign-based operators of a pretend information web site used Claude to provide US midterm protection favoring Russian, Iranian, and North Korean narratives, and different actors used Claude to provide fabricated information designed to draw American audiences and generate promoting income. Anthropic mentioned it banned the accounts concerned. Its September 2026 menace intelligence report documented 9 disrupted affect operations and assessed that the majority such content material drew little genuine engagement, partly as a result of Anthropic usually disrupts operations in the course of the manufacturing stage, earlier than materials reaches the general public.
Anthropic mentioned it hopes to publish a extra complete account of how individuals used Claude all through the election cycle within the coming months, and described the midterm work as extending to elections world wide.
