Anthropic rewrote the rules for what people may do with Claude, and the clause everyone noticed first is the one about being cruel to the software.
The company published its annual Usage Policy update on October 8. It takes effect on November 12. Most of the language formalizes restrictions that were already spread across separate sections on elections, fraud, privacy, and disinformation. The revision also adds rules that only make sense because Claude now runs longer jobs with less human supervision, including a requirement that someone who can stop the machine has to be watching when Claude is wired into hardware that moves.
The reason this matters now is that usage policies have become the main place where AI companies commit to something enforceable. Anthropic's September threat intelligence report, which the policy update cites directly, documented state media outlets, government propaganda offices, and commercial firms using Claude to run networks of fake accounts and fabricated news sites. The company is not describing hypothetical risk. It is describing cases it has already seen, and the November 12 rules are the contract it intends to point at when it shuts those accounts down.
The abuse provision is the narrowest piece of the update. Anthropic added a prohibition on "sustained and needless abusive or cruel behavior toward our models." The post says it applies "only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose." Frustration is still allowed. Pushback is allowed. So are dark creative themes, red-teaming, and academic research.
That distinction is doing a lot of work, and it is not clear the line will hold once a moderator has to judge a borderline case. Anthropic says Claude's ability to end hostile conversations, a behavior it shipped in August, will remain the primary enforcement mechanism. The policy language is a backstop, not the tool.
The update also lands a month after reporting that Anthropic had approached religious scholars, in some cases arguing that Claude could be conscious or possess a soul. Reading the abuse clause next to that reporting makes it hard to separate the product decision from the moral one. The company frames it as user conduct. It reads at least partly like an institution describing how it wants its models treated.
Deceptive campaigns get their own section
The sharpest change is structural. Anthropic consolidated its scattered prohibitions on deception into a single section titled "Do Not Engage in Deceptive Campaigns or Artificial Activity." It covers efforts to obscure who is behind a message, amplification through fake accounts or posts, and building the infrastructure that runs an influence operation.
Two details stand out. The section applies to commercial deception, not just political deception, so a marketing firm running thousands of synthetic reviewers is covered alongside a propaganda office. And it reaches people who build the tooling rather than only the people who pull the trigger on a campaign. Anthropic did not name the operators in its threat report, so there is no way to check from the outside how many of them were enterprise customers rather than burner accounts.
Elections rules narrowed in both directions
The elections section is now titled "Do Not Undermine Democratic Processes," a narrower frame than the old language. It bans using Claude to deceive voters, impersonate candidates or election officials, and suppress turnout.
Anthropic also removed a blanket prohibition on personalized vote and campaign targeting. The company says the original rule swept up legitimate civic work, naming nonprofits using Claude to write voter information in other languages and election officials sending ballot cure notices. Targeting built on deception or misused personal data stays prohibited under the other sections.
That is a real loosening, and it is worth flagging as such. Anthropic is trusting users to keep persuasion separate from fraud, which is a harder distinction to police than a flat ban. If the deceptive campaigns section catches everything the company wants caught, the change is sensible. If it does not, the company just widened the gap between the letter of the policy and what it can detect.
Weapons software and autonomous hardware
Anthropic's weapons language always covered the hardware. The update makes clear it covers the parts that make hardware work. Guidance and control software for weapons is now named explicitly, as are actions like arming drones and other autonomous vehicles. Anthropic says this reflects how it already enforced the prior policy rather than a new posture.
The same gap-closing logic shows up in the hardware section. Anthropic shipped a Model Hardware Standard research preview, and the policy now attaches requirements to it: a qualified operator has to be able to observe the equipment and stop it, and the equipment has to be able to hold a safe state if Claude loses its connection. That is a sane requirement, and it is also the kind of thing that is easy to write and expensive to verify.
Surveillance language gets precise, not softer
The surveillance and criminal justice section was rewritten to state exactly what Anthropic prohibits. Tracking people without consent is banned, including tracking done by analyzing data collected earlier. Claude cannot be used to decide or recommend who gets investigated, arrested, or charged.
Anthropic also listed what stays permitted: fraud monitoring, content moderation, journalism, and legal research. Naming the exceptions is more useful than most vendors manage, and it gives defense contractors and law enforcement agencies a clearer map of what they can actually buy.
High-risk use rules unchanged, just answerable
Where Claude's output affects health, legal rights, finances, livelihood, or access to essential services, the policy requires a qualified human who can review and override the recommendation, and it requires telling the affected person that AI was involved. Those requirements are not new. What is new is that the rewritten section lists which kinds of recommendations are covered, because customers kept asking.
The Supported Regions policy got a similar clarity pass. Anthropic now spells out that access is prohibited for people physically located in an unsupported region, entities incorporated or headquartered there, and entities majority-owned or controlled by persons in those regions.
What to watch
Every section of this update claims to be a clarification. Some of it is. The abuse clause, the hardware requirements, and the explicit ban on targeting tools are new commitments, and the November 12 date is the moment they become real.
The enforcement mechanism is the thing to track. Anthropic has said Claude ending an abusive conversation will do most of the work on that clause, which means the policy will be tested by the cases where the company decides not to enforce it. The same question applies to the deceptive campaigns section and the loosened vote targeting rule. A policy is only as good as the first hard call someone at the company has to make at 2am, and nobody outside has seen that call yet.
More AI coverage is on our AI desk. Anthropic's full announcement is on its newsroom page, and TechCrunch reported the abuse provision first.
