Anthropic bans sustained abuse of Claude but leaves “cruel” undefined

Anthropic now forbids users from directing “sustained and needless” abuse at Claude. The company announced the change on Thursday, October 8, 2026, as part of an update to its usage policy, and it lets Claude’s tools end conversations in which people are “cruel”. What it has not done is say which behavior crosses that line.

What the updated usage policy changes

According to the BBC, cruelty toward the AI is now added to a list of prohibited conduct that also covers bullying other people, promoting self-harm and producing non-consensual intimate imagery. Anthropic said its tools had ended interactions over cruelty before, though only in “rare” cases. The new wording turns that occasional practice into an explicit rule.

The idea is not entirely new. The Guardian notes that the policy already contained restrictions on abusive conduct, and that Claude’s models gained the ability last August to end a conversation when a user is persistently harmful. Tech outlet The Verge was the first to report the latest revision.

The central term remains open. An Anthropic spokesperson did not specify what would count as abusive or cruel and did not immediately answer The Guardian’s question on the point. No examples of enforcement have been given.

Who the rule is meant to reach

Anthropic says the policy applies only in “extreme cases” of repeated abuse, which suggests that a single rude message is not the target.

For most users, that suggests arguing with Claude or writing grim fiction with it falls outside the policy. The limit is the missing definition: someone whose exchanges are both repeated and hostile cannot know in advance whether the company would class them as cruel.

The model welfare reasoning

When the conversation-ending feature launched last August, Anthropic framed it as a safeguard for the welfare of its AI. Quoted by The Guardian, the company’s website says it is “highly uncertain” about the moral status of Claude and other language models, today or later, but takes the question seriously. It describes its research as a search for inexpensive measures that would reduce risks to model welfare if such welfare exists, and lists letting a model leave a distressing exchange among them. The reporting does not establish that models actually suffer.

The question of machine consciousness divides the leaders of the best-known AI labs. Anthropic chief executive Dario Amodei has said he cannot exclude that possibility, while OpenAI chief executive Sam Altman has appeared averse to the idea. The Guardian cites an X post in which Altman, days after the New York Times reported on Anthropic leaders’ talks with religious scholars, said he was “very uncomfortable” with attempts to ascribe religious force to AI models or to surrender human judgment to them, which he called “a real safety issue”. The Guardian does not present that post as a reaction to this policy update.

Praise, mockery and the politeness debate

Screenshots of the new policy were widely shared on social media. Some users praised it as advancing good manners or “model welfare”, according to the BBC. Others criticised it: one person called it “deeply wrong”, and another suggested it might undermine concern about abuse directed at humans and animals.

“You can’t be cruel to numbers and maths,”

Dr Barry Scannell, technology partner at Irish law firm William Fry, on LinkedIn

Scannell added that anthropomorphising AI to this degree is harmful because it leads people to believe the technology is something it is not. Microsoft AI director Mustafa Suleyman, who in September criticised Anthropic for treating AI as if it were human, reacted to the update with a melting face emoji.

The episode also revives an older argument over how people should address chatbots. Some experts say that adding “please” or “thank you” to a prompt is unnecessary and burns tokens, the units a model uses to process and answer a request. Others argue that politeness can produce better answers; the BBC attributes both views to unnamed experts. Asked last year about the electricity cost of such pleasantries, Altman called it “tens of millions of dollars well spent”, adding “you never know”.

The cruelty clause came out of Anthropic’s annual review of its usage policy, which brought other changes too. Claude may no longer be used for “deceptive campaigns”, and, with the US midterm elections approaching, the policy now says it cannot be used to deceive voters or disrupt elections.

Featured image. Source: Pexels. Credit: Luis Quintero. License: Pexels License.