Anthropic's Claude policy rewrite bars building surveillance tools and weapons software

The sentence a public-safety software vendor will read twice in Anthropic's new usage policy is this one: "Claude cannot be used to decide or recommend who to investigate, arrest, or charge in a law enforcement or criminal justice process," the company wrote ^[10]. A lead-ranking feature that leaves the final call to a detective still recommends.

Two of the changes come from what people actually did with Claude ^[8][9]. Anthropic said it saw multiple attempts to use the model for weapons guidance and control software, despite a standing ban on weapons development, so the ban now reaches "software and components that make weapons work, as well as actions like arming drones and other autonomous vehicles" ^[8]. Its last threat intelligence report suggested Claude-powered surveillance was increasingly used to identify and track "political dissidents," and the surveillance bans were made clearer ^[9].

"Tracking people without their consent is prohibited, whether it happens in real time or through analysis of previously collected data," Anthropic wrote ^[10]. A tool that rebuilds someone's movements from stored logs counts the same as a live tracker. So does building one: Claude may not be used to "build or improve tools designed for surveillance" ^[10].

Messaging vendors get one clause in place of "scattered" restrictions, aimed at "efforts to obscure who is behind a message or amplify content through fake accounts or posts" ^[6]. The election section adds impersonating candidates or election officials and trying to suppress turnout ^[7]. A feature that spins up persona accounts and posts at volume is the one to test.

Robotics teams get a design requirement. Claude "connected to hardware that takes autonomous physical actions and which might be capable of causing injury" needs a "qualified operator" able to "observe the equipment and stop it if needed" ^[12]. The Verge noted it is unclear whether that operator must be human ^[12].

The rules may be modified for "contracts with certain governmental customers" when Anthropic judges the contract restrictions and safeguards adequate ^[11]. The company has previously contracted with the US military ^[11]. The Verge's account lists health and financial uses among the new high-risk categories but does not describe those rules ^[1].

The clause that drew the headline touches builders least. It bars "sustained and needless abusive or cruel behavior" toward Claude ^[2] and, Anthropic wrote, "does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research" ^[5]. Ending the conversation, a power Claude got last August under Anthropic's "model welfare" research ^[3], remains the "primary enforcement mechanism" ^[4]. "We don't know if the models are conscious," chief executive Dario Amodei said on a podcast in February ^[13].

For Monday's review, sort each feature on two axes. One is whether it makes or suggests a call about a specific person: who to track, flag or investigate. The other is whether it acts without a person approving each step, through hardware or by posting at volume. Features that do neither face the general terms, plus the health and finance rules in those sectors. Adding human sign-off to a person-specific feature does not clear the surveillance clauses, since recommending is banned ^[10]. For autonomous features that target no one, the stop-control rule ^[12] or the fake-accounts rule ^[6] applies. Where a feature does both, it is squarely inside the new bans, and the one exit the reported text offers is a government contract Anthropic judges adequately safeguarded ^[11].