Anthropic's revised usage policy requires a human in the loop for high-risk Claude use

If the clarifications are what Anthropic says they are, a Claude deployment that complied before Thursday still complies ^[1]. The high-risk rule is described as a new requirement. The Deep View gives its scope as people's "health, legal rights, finances, livelihood, or access to essential services" ^[6]. On that wording, hiring tools and benefits triage are in scope alongside lending and clinical work.

The abuse clause drew the headlines ^[2], but the high-risk rule is the one that will change system diagrams. A long-running agent is useful because it chains tool calls without a person approving each one. In my view the workable design puts the person at the point of effect. That is the call that moves money, updates an eligibility record or sends a decision to the person it concerns. Planning and retrieval before that call can still run unattended. The approval becomes a queue with a timeout. The agent's state has to survive the wait, so the agent needs durable checkpoints.

The newsletter does not quote the policy text, give an effective date or say what qualifies as a human in the loop. Those details decide the build. A reviewer who approves each consequential action before it executes is a different control from one who samples a batch afterwards. Only Anthropic's wording settles which one it means.

The surveillance section needs the same close read ^[5]. A feature that ranks people for investigators produces the exact output that section now names. That holds whether or not the product calls itself surveillance.

The consolidated deceptive-campaign section ^[3] is aimed at the pattern OpenAI described the same day ^[7]. OpenAI said the Iranian operation it ended created personas of seven journalists to pitch stories to small and medium-sized outlets ^[8]. The Russian one drew people in Latin America into running an on-the-ground think tank with fake documents and audio scripts, according to OpenAI ^[9]. "What is most striking about these operations is that they closely resembled complex influence operations of the pre-AI age, but used AI to make some of the workflows easier," OpenAI said in its blog post ^[10].