The Cruelty Clause Is Narrow
Anthropic's 8 October update adds a ban on sustained and needless abusive or cruel behaviour toward its models, and it says the rule applies only in extreme cases where a user is cruel with no discernible purpose.
It does not cover what Anthropic calls common user frustration, pushback, dark creative themes and model testing or research. Enforcement stays with Claude's ability to end abusive conversations on Claude.ai and Claude Code. The Register notes it is unclear how Anthropic will tell frustration from gratuitous cruelty and that there is no established evidence Claude experiences distress.
What Changes On 12 November
The updated policy takes effect on 12 November 2026, and Anthropic says most of the updates clarify existing rules.
| Area | What the update says | Who feels it |
|---|---|---|
| Deceptive activity (new section) | Bans deceptive political or commercial activity, including hiding who is behind a message, amplifying content through fake accounts and building influence-operation tools | Marketing, growth and agency teams |
| Elections | Refocused on deceiving voters or disrupting elections; the blanket ban on personalised vote and campaign targeting is removed | Civic and campaign tools |
| Weapons | Covers software and components that make weapons work, and arming drones or other autonomous vehicles | Defence and robotics suppliers |
| Surveillance | No tracking people without consent; Claude cannot decide or recommend who to investigate, arrest or charge; no building surveillance tools | Security, public-sector and investigation tools |
| High-risk uses | A qualified human can review and change recommendations and the affected person is told AI was used; requirements unchanged but rewritten | Health, legal, finance, livelihood and public-service workflows |
| Physical action | A qualified operator can observe and stop the equipment, which must hold a safe state if Claude disconnects | Robotics and industrial automation |
| Supported regions | Bars people in unsupported regions, entities incorporated or headquartered there and entities majority-owned or controlled from there | Any company with cross-border ownership |
The post summarises the changes and does not reproduce the full policy, so read the policy itself for the list of covered recommendations.
Where A Business Has To Change Something
Four clauses ask a business to change how it works.
Human review. Where Claude affects health, legal rights, finances, livelihood or access to essential services, a qualified human must be able to review and change its recommendations, and the affected person must be told that AI was used. Anthropic says these requirements are unchanged but rewritten, and the full policy lists which recommendations are covered.
Physical action. If Claude is connected to equipment that takes autonomous physical actions and could cause injury, a qualified operator must be able to observe and stop it, and the equipment must hold a safe state if Claude is disconnected. Anthropic ties this to its Model Hardware Standard.
Fake accounts and ownership. Amplifying content through fake accounts or posts sits in the new deceptive-activity section, which covers commercial activity as well as political. The supported-regions rule reaches entities majority-owned or controlled by persons or entities in unsupported regions, so a company registered in a supported country can still be caught by who owns it.
A Checklist Before 12 November
List every workflow that sends Claude output into a decision about a person, and record who reviews it, what authority they have to change it and where the person is told AI was used.
List every place Claude can act on equipment or a live system, and test the stop control and the disconnect state. Review marketing and growth automation for patterns that look like fake-account amplification, and ask legal to check your ownership chain against the supported-regions page.
For readers in Europe, the human-oversight and notice themes will sound familiar from the EU AI Act. Article 14 asks that high-risk systems can be effectively overseen by people, with a stop button or similar procedure that brings the system to a halt in a safe state, and Article 26 asks deployers to assign oversight to people with competence, training and authority and to inform the people subject to a high-risk system. The texts differ, so do not treat compliance with one as proof of the other.
Read next: Your AI Safety Promise Can Now Cost You the Contract | The Pentagon Lost One Anthropic Case and Won the Next
