Anthropic updates Claude rules on physical safety and deceptive campaigns

New text, a later start date

Anthropic published an updated Usage Policy on October 8, setting November 12 as its effective date. The company says much of the revision makes existing restrictions more explicit as Claude handles longer and more independent tasks. Some provisions do change, including controls for physical equipment and the treatment of personalised campaign targeting.

This is a set of conditions for using Anthropic's services, not a new law or evidence that a model will reliably follow every rule. Customers still need to assess their own uses, and the policy does not replace applicable legal requirements.

Deception rules move beyond elections

The update gathers restrictions on deceptive campaigns into one section covering political and commercial activity. The text addresses fake personas, concealed sponsorship and networks of supposedly independent outlets. It also prohibits seeding search engines or AI answer sources with material that misrepresents its origin or independence.

The earlier policy expressly banned personalised vote or campaign targeting based on individual profiles or data. Anthropic is removing that blanket prohibition. It gives multilingual voter information and ballot-correction notices as examples of legitimate civic work that the old wording could capture. Deceptive targeting and misuse of personal data remain prohibited under other sections.

The distinction matters: changing the targeting rule is not permission to impersonate a candidate, suppress turnout or disguise an automated campaign.

A model cannot be its own physical safety limit

The revised high-risk section requires a qualified operator to observe autonomous equipment and be able to stop it. Equipment must stop or hold a safe state if the operator intervenes or the connection to Anthropic is lost. Bounds such as speed or force must be enforced by the equipment or a controller independent of the model's output.

This follows Anthropic's August research preview of the Model Hardware Standard, an interface intended to connect agents with programmable laboratory and manufacturing instruments. That earlier proposal standardises communication with devices; it is not proof that all connected equipment is safe.

In practical terms, a model deciding that an action looks acceptable is different from an independent mechanism preventing an unsafe action. The policy makes that separation explicit.

Review the workflow, not just the prompt

Both the previous and revised policies call for qualified human involvement and disclosure in high-risk advice or decisions. The new text identifies covered recommendations more precisely and includes narrow exceptions for specified decisions wholly favourable to an individual.

For teams using Claude, the useful review is therefore about what the system can actually decide or do, who can intervene and what happens when a connection fails. Enforcement consistency and the effectiveness of those controls in real deployments remain open questions; the announcement does not supply an independent assessment.