Anthropic is lifting its blanket ban on personalized targeting for votes and campaigns, while introducing safeguards for Claude operating equipment autonomously that could injure people. The revised Usage Policy takes effect November 12, with deceptive voter targeting and misuse of personal data still prohibited. Anthropic says the targeting restriction also blocked legitimate activities, including nonprofits preparing voter information in different languages and officials notifying voters about ballots needing correction.
For potentially dangerous equipment connected to its models, Anthropic now requires a qualified operator who can watch the hardware and intervene to stop it. The hardware must also maintain a safe condition if its connection to Claude breaks. Existing requirements for consequential uses remain: a qualified person must have power to review and change Claude's recommendations, and affected individuals must receive disclosure that AI was involved.
Read more: Anthropic Says Claude Found a CRISPR-Like Enzyme System
Anthropic is also introducing a rule against repeated cruelty or abuse directed at its models when it serves no discernible purpose. The company says this applies to extreme behavior, while ordinary frustration, challenges to responses, dark fiction, testing and research remain outside its scope. Claude's existing ability to terminate rare, persistently abusive conversations in Claude.ai and Claude Code will remain the main way Anthropic enforces that restriction, according to the announcement.
The election changes narrow the focus to voter deception and interference, including attempts to reduce turnout or impersonate election officials. A separate section consolidates restrictions on deceptive campaigns across commercial and political activity, covering concealed message sponsors, fake-account amplification and infrastructure for influence operations. Anthropic says it has observed government propaganda offices, commercial businesses and state media using Claude for fabricated news websites and networks of bogus accounts.
Anthropic says its weapons and surveillance revisions spell out restrictions it already enforces, rather than changing enforcement practices. Weapons restrictions explicitly extend to guidance software, functional components and equipping autonomous vehicles with weapons. The surveillance rules prohibit tracking without consent, including through historical data, and bar Claude from selecting or recommending people for investigation, arrest or criminal charges.
The company continues to permit consent-based tracking, including fraud monitoring, alongside journalism, legal research and content moderation. Its regional restrictions cover users physically present in unsupported regions and organizations based, incorporated, majority-owned or controlled there.
Earlier this year, Anthropic separately revised its Responsible Scaling Policy, replacing unconditional commitments to pause dangerous development with conditions tied to its competitive position and assessment of catastrophic risks. That revision introduced Frontier Safety Roadmaps and Risk Reports on a three-to-six-month schedule.
