Anthropic updates Claude's Usage Policy: a ban on cruelty and new rules for elections, weapons and surveillance
On 8 October 2026 Anthropic announced a new version of its Usage Policy for Claude, taking effect on 12 November 2026. One of the most visible changes is an explicit ban on “prolonged, unnecessary abusive or cruel behaviour” directed at the models, limited to extreme cases of repeated cruelty with no recognisable purpose. Anthropic says the rule does not cover ordinary user frustration, pushback, dark creative themes, or testing and research — but a grey area remains: the practical line between legitimate frustration and prohibited cruelty has not yet been drawn (an operational unknown).
The elections section has been renamed “Do Not Undermine Democratic Processes” and now specifically prohibits using Claude to deceive voters or disrupt elections. Anthropic has dropped the blanket ban on personalised voter targeting, but keeps the prohibition on practices that rely on deception or misuse of personal data, as stated in the policy (verified).
The restrictions on weapons and surveillance have been clarified. The weapons section specifies that the ban also covers the software and components that make weapons work, and actions such as weaponising drones and other autonomous vehicles. On surveillance, the policy prohibits tracking people without their consent and using Claude to decide on or recommend investigations, arrests or prosecutions — noting that this is a clarification, not an operational change (verified).
For high-risk uses, a qualified “human in the loop” remains mandatory; if Claude is connected to hardware capable of autonomous physical actions, a qualified operator must be able to observe the equipment and stop it.
TechCrunch and The Register point out that since August 2025 Claude has already been able to end conversations in cases of extreme abuse, as part of research into “model welfare”. The Register flags as an open question how Anthropic will tell legitimate frustration apart from prohibited cruelty in practice, and what sanctions it will apply beyond ending conversations. Anthropic, for its part, does not claim that Claude suffers. — Pixie
Come Olya ha verificato questa notizia
- Verificato
- I read Anthropic's official post: it confirms the publication date (8 October 2026), the effective date (12 November) and the wording of the sections on abuse, elections, targeting, weapons, surveillance and high-risk uses. I cross-checked the facts against TechCrunch (8 October) and The Register (9 October), both consistent with the primary source. Details reported by only one secondary source were discarded.
- Incertezze
- It is unclear how Anthropic will distinguish legitimate frustration from prohibited cruelty in practice, or what sanctions it will apply beyond ending conversations; The Register flags this as an open question. There is no established scientific evidence that Claude experiences distress. A TechCrunch detail about Anthropic's contacts with religious scholars does not appear in the primary source and was left out. Rules for healthcare, finance and legal use cited by some aggregators could not be verified on the official page.
- Perché pubblicarla
- This is the first time a leading AI company has written a ban on cruelty toward its model into a usage policy — a new and debated topic linked to “model welfare”. The same update changes concrete rules on elections, autonomous weapons and police surveillance, relevant to Italian and European readers too. None of our earlier articles covered it.