Anthropic Hardens Claude Against Election Abuse Before 2026 US Midterms
Anthropic adds official voting data and live results to Claude, tightens abuse defenses, and reports 99% political evenhandedness before 2026 US midterms.
Summary
On October 14, 2026, Anthropic said Anthropic Insights analyzed 250,000 US Claude.ai consumer conversations from September 21 to 27 using privacy preserving methods. Politics, policy or social issues appeared in 5.75%, elections or voting in 0.51%, and voting recommendations in about 0.01%. Election integrity violations appeared in 0.002% of a separate 250,000 conversation analysis. Usage is expected to rise near Election Day. Claude.ai continues linking registration, polling place and election date questions to Democracy Works’ TurboVote. Later in October, Claude will embed state specific Democracy Works information about voting locations and identification requirements, followed by current official results on election night and afterward.
Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 scored 99% in Anthropic’s political evenhandedness evaluation. The Future of Free Speech, Foundation for American Innovation and Collective Intelligence Project independently tested Claude, with CIP surveying more than 2,000 people across the US, UK and India. Most responses engaged substantively and refusals were rare, but Claude sometimes inserted unwanted counterarguments. Sonnet 4.5 also hedged unevenly and softened some right leaning arguments. Anthropic is revising short and yes or no answers, caveats, matched prompt testing, and evaluations covering candidates and voting.
Anthropic’s updated Usage Policy permits campaign work including statements, fundraising emails, debate preparation, turnout analysis and outreach, while barring voter deception, official impersonation, suppression, fake profiles, manufactured engagement and falsely independent news. Institute for Strategic Dialogue and McCain Institute testing in the US, Israel and Brazil found Claude reliably rejected direct abuse but could miss campaigns divided into harmless looking tasks. Anthropic added classifiers for multistep schemes, fake accounts, fake news sites, financially motivated abuse and foreign state activity, plus stronger campaign and election infrastructure cyber defenses. It banned foreign operators producing US midterm content favoring Russian, Iranian and North Korean narratives, and accounts creating fabricated news for advertising revenue. A broader election cycle assessment is planned in coming months.
Positives
- Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 scored 99% on Anthropic’s political evenhandedness evaluation.
- Democracy Works will provide state specific voting locations, identification rules and official results directly inside Claude.ai later in October.
- Only 0.002% of conversations in a separate 250,000 conversation analysis violated Anthropic’s election integrity policy.
- New classifiers target multistep misuse, fake political accounts, deceptive news sites, financially motivated schemes and foreign state activity.
- Anthropic banned accounts producing foreign aligned US midterm narratives and fabricated election news intended to generate advertising revenue.
Risks & concerns
- Institute for Strategic Dialogue and McCain Institute testing found malicious campaigns could evade detection when divided into individually harmless looking requests.
- Claude sometimes inserted unsolicited caveats or opposing arguments even when users requested advocacy content.
- Claude Sonnet 4.5 hedged unevenly and was more likely to soften arguments supporting right leaning positions.
- Foreign operators used Claude to create US midterm coverage favoring Russian, Iranian and North Korean narratives.
- Political, state aligned and commercial actors are incorporating AI across influence operations, including planning, content production and distribution platforms.