Watchtower

Anthropic updated its Usage Policy on Oct. 8, detailing how its products are allowed to be used. The new policy will go into effect Nov. 12. Many of the changes seem to have been made with increasingly agentic use of Anthropic’s models in mind, as the models become more capable of a wide range of actions, as well as becoming integrated with physical and other electronic systems.

The largest change was the “High-risk Use Case Requirements” section being extensively rewritten. The requirements now explicitly do not apply to using Anthropic products “to assist in workflows that are related to, but do not actually provide, High-risk AI Recommendations,” such as internal drafting or analysis or carrying out a decision a person has already made.¹ There’s also an exception to the human-in-the-loop review requirement for decisions “wholly favorable to the individual,” such as a person’s insurance claim being fully paid.

High-risk Areas were updated: Public benefits and services, and Legal status and adjudication were added, while “Media or professional journalistic content” was removed as a High-risk Area.

A “High-risk Physical Actions” subsection was added under the High-risk Use Case Requirements section. When connecting Anthropic models to hardware that can take physical actions, users should assess what could happen if something went wrong and “the degree of control the user maintains relative to the autonomy the model has.” Requirements apply when outputs are acted upon without human approval by hardware that can do one of six listed things, such as moving through a shared physical space. These requirements don’t apply to consumer devices operating within a manufacturer’s built-in safety limits where the worst outcome is “discomfort or inconvenience,” handling non-hazardous materials inside enclosed instruments, or generating plans for a qualified person to review before they are executed on equipment.

The bar for enforcing the Usage Policy lowered, with “If we learn that you have violated” becoming “If we suspect that you may have violated.”² Correspondingly, new responses to a potential violation were added, including warnings or limiting access to Anthropic products and services.

The “Do Not Compromise Computer or Network Systems” section was clarified to explicitly allow for red-teaming and other security testing done with permission.³ It now reads “Do Not Compromise Computer or Network Systems Without Authorization,” and that new language was added to most of the section’s prohibitions. 

The “Do Not Develop or Design Weapons” section was extended.⁴ It newly prohibits developing or operating software or components for weapons, illegally selling or transferring weapons, and modifying or testing biological or chemical agents to optimize traits such as their lethality or transmissibility. Overall, the section became more specific and wide-ranging, aside from loosening its harm standard from “designed to cause harm to or loss of human life” to “designed to deliver lethal or destructive force.”

Anthropic newly prohibits “sustained and needless abusive or cruel behavior toward [Anthropic’s] models.”⁵

A new section was added to the Usage Policy: “Do Not Engage in Deceptive Campaigns or Artificial Activity.”⁶ This section prohibits creating or operating fake personas or impersonating real people in order to mislead people. It prohibits misrepresenting sponsorship of content, distributing content through networks of fake outlets or accounts, building tools designed to assist with deceptive campaigns, as well as soliciting funding for any of the above as a product or service.

The “Do Not Undermine Democratic Processes” section was edited. It no longer prohibits engaging in personalized vote or campaign targeting, while newly prohibiting the creation of content that falsely appears to come from or on behalf of a real candidate, official, election authority, or news outlet.⁷

The “Do Not Use for Criminal Justice, Censorship, Surveillance, or Prohibited Law Enforcement Purposes” section was extended.⁸ The update newly prohibits suggesting decisions in law enforcement or criminal justice processes rather than just making them. Instead of just prohibiting nonconsensual targeting/tracking of a person, it now prohibits tracking their property. It newly prohibits building or improving surveillance capabilities of tools designed for surveillance prohibited in this section, and prohibits identifying people from anonymous or pseudonymous activity in order to harass, intimidate, surveil, or harm them. A final paragraph clarifies this section doesn’t prohibit analysis/tracking that individuals have agreed to, nor does it prohibit aggregated/anonymized analysis.

This is not an exhaustive list, as the Usage Policy was extensively updated.

A diff of the changes can be found below: