Anthropic’s revised rules for using Claude took effect Thursday, spelling out when employers, insurers and other organizations need a qualified person to review a recommendation before acting on it. The policy covers Claude’s consumer apps, Claude Code and services that use the company’s models through its API or authorized providers.
Anthropic announced the revisions Oct. 8. It says the basic review and disclosure obligations for high-risk advice were already in place. The new text names the decisions that trigger them and adds controls for some uses of Claude with equipment that can injure people.
Which recommendations need a reviewer
Screening or ranking job applicants is on the list, as are setting a worker’s pay, deciding whether someone qualifies for a loan and recommending a medication change. When Claude is used to make a covered recommendation, a person with relevant training or experience must review it before the advice goes out or the decision takes effect. That reviewer needs authority to change the recommendation and a license where the law requires one.
The person receiving the advice or affected by the decision must also be told AI was used. The disclosure can omit the name of the model. Consumer-facing chatbots have a separate disclosure rule. They must tell people they are interacting with AI at the start of the session or in the interface.
Internal research and drafting fall outside the high-risk review requirement when the Claude output stays separate from the recommendation ultimately delivered. An HR team using Claude to summarize general workplace guidance is in a different position from one using it to rank individual candidates. The policy also gives a limited exception from prior human review for certain fully favorable decisions, such as paying an insurance claim, where the law allows it.
What the policy adds for machines and chat sessions
When Claude output controls certain equipment capable of causing injury, the policy requires a qualified person who can watch the equipment and stop it. The machine must stop or remain in a safe state if that person intervenes or it loses its connection to Anthropic. Safety limits have to work independently of Claude’s output.
The update also expressly prohibits repeated, purposeless cruelty toward the model. As The Verge reported when the policy was announced, Anthropic had already allowed Claude to end some conversations involving persistent abuse. The company says ending those interactions remains its main enforcement method and that ordinary frustration, creative work with dark themes, and model testing fall outside this prohibition.
