Anthropic has announced a revision to its usage policy that becomes effective on November 12, 2026. The most consequential question is not whether Claude can produce an answer, but who remains responsible when that answer influences a person or moves equipment.
The change combines revised prohibitions with more explicit operating requirements. Its dates also matter: an announced future policy should not be described as a rule already in force. Developers have a document to examine before the new version takes effect.
A clearer boundary around influence and misuse
Anthropic’s announcement says much of the revision clarifies existing restrictions rather than creating entirely new bans. It brings deceptive political and commercial campaigns into a common framework, including misleading networks of accounts or websites that disguise their origin.
The election rules focus on deception and interference with participation or electoral infrastructure. The company removes a blanket restriction on personalized political campaigning, while retaining prohibitions on misleading voters and violating privacy. Removing one broad rule is therefore not permission to use the system for covert manipulation.
Anthropic also narrows how it describes prohibited surveillance and weapons-related activity. It distinguishes consent-based monitoring and legitimate research from tracking people without consent or assisting prohibited law-enforcement decisions. These are statements of the company’s permitted-use boundaries, not a finding that all deployments already meet them.
The distinction is practical. A policy can identify conduct that should not occur; it cannot, by its publication alone, demonstrate how reliably that conduct will be detected or prevented. Evaluation of enforcement remains separate from reading the revised terms.
Human review must include real authority
The policy text requires meaningful review by a qualified person before high-risk recommendations are delivered or used for decisions. That reviewer must have authority to change the output, and affected people must be told that AI was involved.
For high-risk physical actions, a qualified individual must be able to observe operation and stop equipment. The equipment must stop or hold a safe state when that person intervenes or when the connection to Anthropic’s services is lost. Operating limits must be enforced independently of model output.
These requirements describe more than an approval button. Reviewing a recommendation without the competence or authority to alter it would not satisfy the policy’s stated standard. Likewise, an instruction asking a model to behave safely is not the independent controller the physical-action provisions describe.
The document applies across Anthropic’s apps, APIs and integrated services, including end users. It also distinguishes prohibited uses from high-risk uses that can proceed subject to additional controls. That distinction prevents a human reviewer from being presented as a general exception to every prohibition.
Model welfare and physical capability are different questions
The revision also addresses sustained, needless abusive behaviour toward models. Anthropic’s earlier research on ending conversations supplies relevant background: in August 2025, it described giving particular Claude models a last-resort option to end severely abusive interactions.
That research was explicitly exploratory about AI welfare. It did not establish that models are conscious, and its safeguards sought to preserve assistance for people in immediate danger. The new restriction should therefore not be reported as scientific proof of machine feelings.
Physical integration presents another kind of uncertainty. Anthropic’s Model Hardware Standard preview describes a protocol for connecting models to equipment through standardized interfaces. The company reports an early laboratory demonstration with Genentech, while acknowledging situations where human help remains necessary.
A research demonstration is not evidence that arbitrary equipment can be controlled reliably. The preview makes the policy’s operational questions more concrete: which actions are permitted, what independent limits exist, and who can intervene?
The policy is a set of conditions for use, not a universal safety certificate. Its value will be tested in systems where human authority, disclosure and independent safeguards exist in practice rather than only in documentation.










