Anthropic's new Claude policy requires a qualified operator who can stop the machine

The Verge reported this week that Anthropic has changed its usage policy for Claude. It is the first change in over a year. The clause that drew the headlines prohibits "sustained and needless abusive or cruel behavior" toward the model. If you build on Claude, the clauses further down matter more. They expand the weapons ban and make the surveillance ban clearer. They gather scattered restrictions into one rule against deceptive political and commercial campaigns. They also add a requirement that a "qualified operator" can observe and stop any hardware Claude controls. These rules set what may be built, by whom and under whose supervision.

The welfare clause is narrow by design

Anthropic wrote that the rule "is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose." It excludes "common versions of user frustration, pushback, dark creative themes, or model testing and research." Ending the conversation remains the "primary enforcement mechanism", a tool Anthropic introduced last August. According to The Verge, Anthropic did not say whether user bans could follow.

The practical cost to a working professional is close to zero. The clause matters as a signal: the vendor now treats how people behave toward its model as a policy matter. Microsoft AI took the opposite line in September. Its code of conduct reads: "We reject the pursuit of legal personhood, or the idea that models might deserve welfare, or be entitled to rights." Two major suppliers now hold opposite positions on the same question. Your team works under the rules of whichever one your organisation licensed.

The rules that change what gets built

Weapons. Anthropic said it had seen multiple attempts to use Claude to develop guidance for weapons and software to control them. The ban now covers "software and components that make weapons work, as well as actions like arming drones and other autonomous vehicles."

Surveillance. Anthropic's last threat intelligence report suggested that people were increasingly using Claude-powered surveillance to identify and track "political dissidents". The new text reads: "Tracking people without their consent is prohibited, whether it happens in real time or through analysis of previously collected data." Claude "cannot be used to decide or recommend who to investigate, arrest, or charge in a law enforcement or criminal justice process." Anthropic also prohibits anyone from using Claude to build or improve surveillance tools.

The carve-out. These rules may be modified for "contracts with certain governmental customers" if, "in Anthropic's judgment", the safeguards are adequate. Anthropic has contracted with the US military before. So the line moves depending on who the customer is, and Anthropic decides where it falls.

Physical systems. When Claude is "connected to hardware that takes autonomous physical actions and which might be capable of causing injury," a "qualified operator must be able to observe the equipment and stop it if needed." The Verge notes that the policy does not say whether that operator must be human.

That last rule is the operator story. A model vendor has written into its terms that someone must be accountable for watching the system and able to halt it. AI does the routine work. You do the thinking, and the policy now expects someone to hold the stop switch.

Your Next Move

1. Audit current builds against the surveillance language this month. List every workflow where your team uses Claude to profile individuals, rank them for scrutiny or recommend who gets investigated. This covers fraud screening, internal investigations and people analytics. Check each one against the new text. Write down your conclusion and who signed it off.

2. Name the qualified operator for each system that acts. If Claude touches anything that moves or acts in the physical world, record who can observe it, how they stop it and what qualifies them. That role carries accountability. Make sure it belongs to a named person with the authority to use it.

3. Track vendor policy as a dependency. Anthropic and Microsoft AI already differ on model welfare. Their usage rules will differ elsewhere too. Note which clauses would block your product if your organisation switched supplier, and make one person responsible for re-reading the policy whenever it changes.