Anthropic's updated Usage Policy spells out explicit surveillance and weapons rules for Claude

Anthropic's account of its surveillance rewrite sets the terms for reading the rest. The company wrote that the rewrite "does not reflect a change in what we enforce in practice: it is just intended to make our policy more explicit." ^[13] For a team already inside the old lines, what changes in that area is the document it is checked against. The company says the weapons language likewise reflects how it enforced the previous policy ^[12].

The skeptic's case is simple: if enforcement has not changed, neither has the build. I think that undersells what explicit wording does in a design review. The surveillance section now names a specific output. The policy says Claude may not be used to decide or recommend who gets investigated, arrested, or charged in law enforcement or criminal justice work ^[7]. An agent that hands an investigator a ranked list of people fits that description, whatever the feature is called. The ban on tracking people without consent also covers analysis of data collected earlier ^[6], so a retrospective analytics job falls inside the line just as a live feed does. Tracking people have agreed to, such as fraud monitoring, stays permitted ^[9].

Commercial teams get the same kind of specificity in the deceptive-campaigns section. It covers obscuring who is behind a message, amplifying content through fake accounts or posts, and building the tools and infrastructure that run influence operations ^[5]. Anthropic tied the section to state media outlets, government propaganda offices and commercial firms it had seen using Claude to run fake-account networks and fabricated news sites ^[3].

Election work moved the other way. With the blanket ban on personalised vote and campaign targeting gone ^[14], the trade-off is that a broad rule that was simple to apply gives way to narrower ones that ask the deployer to judge its own outreach. Targeting that relies on deception or misuse of voters' personal data stays prohibited under the deceptive campaigns, surveillance and privacy sections ^[15].

The controls for Claude acting autonomously in the physical world are the hardest part to plan against from the announcement alone, and the clarified health and finance requirements are in the same position ^[8]. Anthropic gives the reason: Claude has taken on longer and more independent work since the last refresh ^[10]. The post does not say what the controls require. The terms are in the full Usage Policy ^[17].

That leaves a sequencing choice for this quarter. A team that signs off on the summary now may carry a re-review into next quarter if the full text asks for more. A team that waits for counsel to read the full text delays any agent launch that acts in the physical world. Anthropic revises the policy each year ^[1]; the text a team builds to now is the one its agents run under until the next revision.