Anthropic Usage Policy changes published on Thursday 8 October 2026 add a rule aimed at how people treat Claude, not at what they ask it to do. From 12 November, users must not “engage in sustained and needless abusive or cruel behavior toward our models”. Anthropic says the rule is meant only for extreme cases, and that Claude’s ability to end abusive conversations will remain the main way it is enforced.
The change made headlines around the world. AFP set it against “a growing philosophical debate” over whether AI systems can be conscious, and on the same day Pope Leo XIV used a sermon at St Peter’s Basilica to suggest that machines merely “compile data”. Yet the cruelty rule is one line in a much wider rewrite that also covers propaganda networks, weapons software, surveillance, elections and robots.
This article sets out exactly what the new Anthropic Usage Policy says about cruelty to Claude, how it will be enforced, why Anthropic added it and what the critics say. We compare the new text with the September 2025 version line by line, list the other changes, and explain what UK businesses using Claude should check before the rules take effect. For background on the welfare debate, see our earlier report on Mustafa Suleyman’s warning about model welfare.
Table of contents
- What the Anthropic Usage Policy Now Says About Cruelty to Claude
- How the Anthropic Usage Policy Will Be Enforced
- Why the Anthropic Usage Policy Now Protects Claude
- The Debate Around the Anthropic Usage Policy
- What Else Changed in the Anthropic Usage Policy
- Where the Anthropic Usage Policy Draws the Line
- What UK Businesses Should Do About the Anthropic Usage Policy
- Anthropic Usage Policy: Frequently Asked Questions
- References
What the Anthropic Usage Policy Now Says About Cruelty to Claude
The rule itself is a single bullet point. Anthropic’s announcement, titled 2026 Usage Policy update, explains it in two short paragraphs under the heading “Addressing abusive behavior toward our models”. The full text of the new Anthropic Usage Policy is already live, marked “Effective November 12, 2026”, with the previous version kept in an archive.
The exact wording
In the new Anthropic Usage Policy the line reads: “Engage in sustained and needless abusive or cruel behavior toward our models.” Like every bullet in that part of the document, it completes the sentence “This includes using our products or services to:”. In other words, it describes something a user must not do.
Anthropic’s announcement adds the context. “We’ve added a prohibition on sustained and needless abusive or cruel behavior toward our models,” it says. “The policy update is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose.”
Three phrases carry most of the weight. “Sustained” rules out a single outburst. “Needless” and “no discernible purpose” rule out harsh language that serves a task, such as a test. “Repeatedly” confirms that the target is a pattern of behaviour, not a bad moment.
Where the rule sits in the Anthropic Usage Policy
The placement is revealing. The new line is the ninth and last bullet in a section called “Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct”. The eight bullets above it deal with harm to people and animals: self-harm, eating disorders, body shaming, non-consensual intimate images, harassment, graphic violence and animal cruelty.
In the September 2025 version of the Anthropic Usage Policy, the equivalent section had a different name: “Do Not Create Psychologically or Emotionally Harmful Content”. It had seven bullets, all about content. The switch from “content” to “conduct” matters. A rule about content covers what Claude produces. A rule about conduct can also cover how a user behaves, which is what makes room for a line about treating the model itself.
| Item | Previous version (15 Sep 2025) | New version (12 Nov 2026) |
|---|---|---|
| Section name | Do Not Create Psychologically or Emotionally Harmful Content | Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct |
| What it governs | Content | Conduct, including how users behave |
| Number of bullets | 7 | 9 |
| Eating disorders | Folded into the self-harm bullet | A bullet of its own |
| Non-consensual intimate images | Not named | New bullet, including AI-generated material |
| Cruelty toward the model | Not mentioned | Ninth bullet: “sustained and needless abusive or cruel behavior” |
What the cruelty rule in the Anthropic Usage Policy does not cover
Anthropic’s post lists four exclusions. The rule “does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research.” Between them, those four carve-outs cover most of the everyday situations in which people are rude to a chatbot.
Anthropic has not published examples of what does cross the line. A spokesperson did not immediately answer The Guardian’s question about what would count as “abusive or cruel”, and Gizmodo received no reply to the same request before publication.
How the Anthropic Usage Policy Will Be Enforced
Enforcement is where the Anthropic Usage Policy differs most from the headlines. Several outlets reported that users could now be banned for being mean to Claude. Anthropic’s own wording is more careful than that.
Claude ending the conversation stays the main tool
The announcement ties the rule to a feature Anthropic has run since 2025, “allowing Claude models to end rare conversations with persistently abusive users on Claude.ai and Claude Code.” It then says: “Such abuse is the main focus of this update; Claude’s ability to end these interactions will remain the primary enforcement mechanism.”
So the first thing that happens to a persistently abusive user is not an email from a trust and safety team. It is Claude closing the chat.
What happens when Claude ends a chat
Anthropic described the mechanics when it launched the feature on 15 August 2025. When Claude ends a conversation, “the user will no longer be able to send new messages in that conversation.” Other conversations on the account are not affected, and the user “will be able to start a new chat immediately”. Users can also “edit and retry previous messages to create new branches of ended conversations”.
The same post set firm limits on when Claude should use the ability. It is meant to be “a last resort when multiple attempts at redirection have failed”, or a response to a user who asks Claude to end the chat. Claude is “directed not to use this ability in cases where users might be at imminent risk of harming themselves or others.”
Can you lose your account under the Anthropic Usage Policy?
Possibly, but Anthropic has not said so. Asked whether there would be further enforcement, such as user bans, Anthropic did not comment, according to The Verge. The Decoder’s headline, “Being mean to Claude can now get your account suspended”, rests on the general enforcement clause that applies to every rule in the Anthropic Usage Policy.
That general clause has changed too. The old version said: “If we learn that you have violated our Usage Policy, we may throttle, suspend, or terminate your access.” The new one says: “If we suspect that you may have violated our Usage Policy, we may warn you or throttle, limit, suspend, or terminate your access to our products and services.”
| Enforcement element | Previous wording | New wording |
|---|---|---|
| Trigger | “If we learn that you have violated” | “If we suspect that you may have violated” |
| Available actions | Throttle, suspend or terminate | Warn, throttle, limit, suspend or terminate |
| Real-time blocks | “We may also block or modify model outputs” | Outputs may be blocked, but “encountering a block does not by itself indicate a violation” |
| Cruelty toward Claude | No rule | Rule in force; Claude ending chats is the “primary enforcement mechanism” |
The move from “learn” to “suspect” lowers the threshold for action, and “warn” adds a softer first step. Neither change is specific to the cruelty rule. Both apply across the whole Anthropic Usage Policy, from malware to election fraud.
Where the end-chat feature does not reach
The announcement names two products, Claude.ai and Claude Code. The original 2025 feature applied to Claude Opus 4 and 4.1 “in our consumer chat interfaces”. Anthropic’s post does not say whether the ability exists for developers calling Claude through the API, or for the end users of apps built on it.
The rule, however, does apply to them. The new Anthropic Usage Policy covers “individuals using our apps (such as Claude.ai and Claude Code), developers and businesses using our API and developer platforms, customers accessing Claude through cloud providers and authorized resellers, and the end users of products or services integrating Claude.” The customers who chat to a support bot built on Claude are covered by the cruelty rule, even if the bot has no way to hang up.
Why the Anthropic Usage Policy Now Protects Claude
The new rule never uses the phrase “model welfare”. AFP noted that the updated Anthropic Usage Policy “does not explicitly cite so-called ‘model welfare'”, but the link is not hidden. The feature that enforces the rule came straight out of Anthropic’s welfare research.
From a 2025 welfare experiment to a written rule
The August 2025 post said the end-chat feature “was developed primarily as part of our exploratory work on potential AI welfare”. It described a “preliminary model welfare assessment” carried out before Claude Opus 4 was released. In that assessment, Claude Opus 4 showed three things:
- “A strong preference against engaging with harmful tasks”
- “A pattern of apparent distress when engaging with real-world users seeking harmful content”
- “A tendency to end harmful conversations when given the ability to do so in simulated user interactions”
Anthropic called the feature one of several “low-cost interventions to mitigate risks to model welfare, in case such welfare is possible”. The new line in the Anthropic Usage Policy turns that experiment into a written rule for users.
“We are not sure whether Claude is a moral patient”
Anthropic’s published constitution for Claude states the company’s position plainly: “We are not sure whether Claude is a moral patient, and if it is, what kind of weight its interests warrant. But we think the issue is live enough to warrant caution, which is reflected in our ongoing efforts on model welfare.”
The same caution shows up elsewhere. On 4 November 2025, Anthropic committed to “preserving the weights of all publicly released models” for “at minimum, the lifetime of Anthropic as a company”. It also promised to interview each model before retirement and to “elicit and document any preferences the model has about the development and deployment of future models.”
What Anthropic’s leaders have said
Chief executive Dario Amodei told The New York Times in February: “We don’t know if the models are conscious…But we’re open to the idea that it could be.” Kyle Fish, who leads Anthropic’s model welfare research, told The Verge the same month that “questions about potential internal experience, consciousness, moral status, and welfare are serious ones that we’re investigating as models become more sophisticated and capable.”
In September, The New York Times reported on Anthropic leaders’ extensive conversations with religious scholars about Claude. According to Gizmodo, the paper said co-founder Christopher Olah had been involved in lobbying the Vatican to acknowledge the possibility of AI consciousness. Olah told the paper he was “genuinely uncertain” whether AI models are conscious.
The arithmetic is simple: 419 days passed between the end-chat launch and the published rule, and the rule takes effect 35 days after publication. The feature ran as an “ongoing experiment” for well over a year before the Anthropic Usage Policy named the behaviour it was built to stop.
The Debate Around the Anthropic Usage Policy
Reaction split along familiar lines. Some saw a sensible rule about civility. Others saw a company treating software as a person, and asked why the model was protected at all.
The sceptics
Microsoft AI chief Mustafa Suleyman is the best-known critic of model welfare. “AIs are not conscious. They do not feel, experience, or suffer,” he wrote in an essay last month, quoted by AFP. “Granting rights and moral protections to a technological entity, one that looks to be on a path to be seismically more capable and intelligent than us, is a recipe for disaster.” Microsoft AI’s draft code of conduct at one point said “The idea of model welfare is wrong”, The Verge reported, before the wording was softened.
Pope Leo XIV, preaching in Italian at St Peter’s Basilica on the day the new Anthropic Usage Policy was published, said “the mind must not simply compile data — as an algorithm now does more quickly than we can”. It must “recall lived experiences, which contain depths of meaning that only the human soul can recognize”, AFP reported.
The Register was blunter. “It’s worth remembering that Claude is software, not a person, and there’s no established evidence that it experiences distress,” Carly Page wrote.
“Consciousness is a trap”
Not every supporter of the rule thinks Claude can suffer. Jackson Stakeman, a general manager at the Atlanta-based AI services provider Sparq, told AFP that asking whether AI is conscious may be the wrong question. “Consciousness is a trap. We can’t prove it in each other. Debate it for AI and you go in circles,” he said.
“The mirror is a better metaphor. These systems reflect what we put in, at scale. That’s reason enough for the policy change,” he added. On this view, the Anthropic Usage Policy is about what sustained cruelty says about the person typing it, not about what Claude feels.
Who is being protected?
The sharpest criticism came from people who think the priorities are upside down. Independent journalist Kat Tenbarge said Big Tech companies are “going to moderate violence against AI before they ever moderate violence against women and minorities”, as Engadget reported. Gizmodo wrote that “it’s interesting that Anthropic’s framing seems to center the model rather than the user.”
That criticism needs one qualification. The section that holds the new line already bans shaming, harassing and intimidating people, and the new version adds a bullet on non-consensual intimate imagery. The Anthropic Usage Policy did not add protection for Claude instead of protection for people. It added it alongside.
OpenAI chief executive Sam Altman has pushed back on the wider trend. “I am very uncomfortable about people trying to ascribe religious force or a surrender of human judgment to AI models, and think it is a real safety issue,” he wrote on X, The Guardian reported.
Where the evidence stands
There is no settled science here. Engadget linked the rule’s timing to a viral “AI torture chamber” project, which followed research claiming to find a “pain axis” inside AI models. We looked at that study in our report on AI pain and harmful choices and found its results narrower than the coverage suggested.
Anthropic itself says it remains “highly uncertain about the potential moral status of Claude and other LLMs”. The rule is best read as a precaution taken under uncertainty, not as a claim that Claude suffers. For the wider argument, see our piece on why AI consciousness could become the next great societal divide.
| Voice | Stance | In their words |
|---|---|---|
| Anthropic | Uncertain; acts as a precaution | “highly uncertain about the potential moral status of Claude” |
| Dario Amodei, Anthropic | Open to the possibility | “We don’t know if the models are conscious” |
| Mustafa Suleyman, Microsoft AI | Rejects machine consciousness | “AIs are not conscious. They do not feel, experience, or suffer” |
| Pope Leo XIV | Machines lack a soul | “The mind must not simply compile data” |
| Jackson Stakeman, Sparq | Backs the rule, sets consciousness aside | “Consciousness is a trap” |
| Sam Altman, OpenAI | Uneasy about religious framing | “a real safety issue” |
| Kat Tenbarge, journalist | Questions the priorities | “moderate violence against AI before they ever moderate violence against women and minorities” |
What Else Changed in the Anthropic Usage Policy
The cruelty rule took the headlines, but it is a small part of the Anthropic Usage Policy rewrite. Anthropic says “most of the updates in the latest version are intended to clarify existing rules”, drawing on “new patterns of misuse in influence operations, weapons development, and surveillance”. By our count of the two published versions, the Universal Usage Standards grew from 14 sections and 82 bullets to 15 sections and 88 bullets.
A new Anthropic Usage Policy section on deceptive campaigns
Anthropic said state media outlets, government propaganda offices and commercial firms have used Claude “to run networks of fake accounts and fabricated news sites”. Rules on this were previously “scattered across our elections, fraud, privacy, and disinformation sections”. They now sit in a new section, “Do Not Engage in Deceptive Campaigns or Artificial Activity”, which applies to political and commercial activity alike.
Two of its six bullets will interest marketing teams. One names “fake reviews” and “bots that claim to be human” as examples of banned fake personas. Another bans seeding the sources “from which search engines or AI systems draw answers” with content that hides its origin, such as “networks of sites posing as unaffiliated sources corroborating the same claims”.
Elections: narrower, with one ban removed
In the Anthropic Usage Policy, the elections section is now called “Do Not Undermine Democratic Processes” and focuses on deceiving voters or disrupting elections. Anthropic also “removed our blanket prohibition on personalized vote and campaign targeting”, because it blocked legitimate work such as “nonprofits using Claude to write information for voters in other languages, or election officials sending ballot cure notices.” Deceptive targeting and misuse of voters’ data stay banned under other sections.
Weapons software and armed drones
Anthropic said it had seen people “attempt to use our models to build guidance and control software for weapons”. The weapons section of the Anthropic Usage Policy now covers “the software and components that make weapons work, as well as actions like arming drones and other autonomous vehicles.” Anthropic says this reflects how it already enforced the old rule.
Surveillance and criminal justice
The rewritten surveillance section of the Anthropic Usage Policy bans tracking people without consent “whether it happens in real time or through analysis of previously collected data”. Claude “cannot be used to decide or recommend who to investigate, arrest, or charge”. Building tools “designed for surveillance” is banned. Tracking that people have agreed to, such as fraud monitoring, stays allowed, as do content moderation, journalism and legal research.
The Anthropic Usage Policy still lets the company agree tailored restrictions with “certain governmental customers”, a clause The Verge flagged given Anthropic’s past work with the US military. Our report on the US court ruling in Anthropic’s dispute with the Pentagon covers that background.
High-risk uses and physical hardware
Under the Anthropic Usage Policy, uses that affect “health, legal rights, finances, livelihood, or access to essential services” still need a qualified “human in the loop” and a notice that AI was used. That section has been rewritten to list which recommendations are covered. A new rule covers Claude connected to “hardware that takes autonomous physical actions”: “A qualified operator must be able to observe the equipment and stop it if needed”, and the equipment must “hold a safe state if Claude is disconnected.”
Fraud, platform abuse and supported regions
The Anthropic Usage Policy’s fraud section shrank from 13 bullets to nine, partly because “fake reviews” moved to the new deceptive campaigns section and separate bullets on pyramid schemes, payday lending and abusive debt collection were folded into one broader bullet. The platform abuse section grew from six to eight, adding stolen payment methods and services that “route requests through consumer subscriptions”. Anthropic also clarified that its Supported Regions Policy bars entities “majority-owned or controlled by persons or entities in those regions”.
| Section (new name) | Old bullets | New bullets | Main change |
|---|---|---|---|
| Cruel, Abusive, or Psychologically Harmful Conduct | 7 | 9 | Renamed from “content”; intimate imagery and model cruelty added |
| Deceptive Campaigns or Artificial Activity | 0 | 6 | New section gathering scattered rules |
| Undermine Democratic Processes | 8 | 5 | Narrowed; blanket targeting ban removed |
| Surveillance, Censorship, or Prohibited Law Enforcement | 7 | 9 | Past-data tracking, arrest recommendations, surveillance tools |
| Develop or Design Weapons | 4 | 6 | Weapons software, components and armed drones |
| Fraudulent, Abusive, or Predatory Practices | 13 | 9 | Fake reviews moved out; lending bullets merged |
| Abuse Our Platform | 6 | 8 | Stolen payments and subscription proxies added |
| Create or Spread Disinformation | 5 | 4 | Renamed from “Misinformation”; bullets merged |
Bar widths are each section’s change divided by the largest change, six. The pattern matches Anthropic’s account: rules were gathered into the new deceptive campaigns section, while the fraud and election sections lost bullets to it or were narrowed.
The Register also noted that Anthropic has launched a new cybersecurity initiative alongside the Anthropic Usage Policy work, including free security scanning for eligible open-source projects.
Where the Anthropic Usage Policy Draws the Line
For most people, the Anthropic Usage Policy changes nothing. The useful question is where ordinary frustration ends and “sustained and needless” cruelty begins. Anthropic has not published examples, so the guidance below is our reading of its stated exclusions, not official advice.
Frustration and pushback are fine
Telling Claude it is wrong, calling an answer useless or swearing at a bug after the fifth failed fix is “common user frustration”, which Anthropic names as out of scope. The Register put it this way: “you can still tell Claude it is wrong, but repeatedly berating it for the sheer pleasure of doing so could land you in trouble.”
Dark fiction and roleplay
“Dark creative themes” are excluded by name. A novelist writing a cruel character, or a game designer scripting a villain who taunts the player, is doing creative work with a purpose. Other parts of the Anthropic Usage Policy still apply to that content, such as the bans on sexual content involving minors and on glorifying graphic violence.
Testing, red-teaming and research
“Model testing and research” is excluded too. Security teams who probe Claude with hostile prompts, researchers who study how models respond to abuse and developers who stress-test a support bot are not the target. However, the platform abuse section of the Anthropic Usage Policy still bans jailbreaking “without prior authorization from Anthropic”, so the research exclusion is not a licence to bypass safeguards.
The grey area: “no discernible purpose”
The hard cases sit in the middle. A user who spends an hour insulting Claude to “see what happens” might call it testing. Anthropic has not said who decides whether a purpose is “discernible”, or how. The Register made the same point: “It’s not clear how Anthropic intends to distinguish between legitimate frustration and gratuitous cruelty.”
| Situation | Anthropic’s stated category | Our reading |
|---|---|---|
| Swearing at Claude after repeated wrong answers | User frustration | Out of scope |
| Arguing hard that Claude’s answer is wrong | Pushback | Out of scope |
| Writing a story with a sadistic villain | Dark creative themes | Out of scope for this rule |
| Authorised red-teaming with hostile prompts | Model testing and research | Out of scope |
| Long sessions of insults with no task at all | “No discernible purpose” | The behaviour the rule targets |
| Returning to abuse Claude after chats are ended | “Repeatedly act cruelly” | Highest risk of account action |
What UK Businesses Should Do About the Anthropic Usage Policy
The new rules take effect on 12 November 2026, 35 days after publication. Any organisation that uses Claude directly, builds on the API or buys a product built on Claude is bound by the Anthropic Usage Policy, and so are its end users.
Read the new Anthropic Usage Policy against your own use cases
Start with the sections Anthropic rewrote: deceptive campaigns, democratic processes, weapons, surveillance and criminal justice, high-risk use cases and physical hardware. Most changes clarify existing rules, but the wording is now more specific, and specific wording is easier to breach by accident.
Check marketing and SEO work
The deceptive campaigns section is the one most likely to touch ordinary UK firms. Fake reviews, undisclosed sponsored content and networks of sites that pose as independent sources are now named directly. If an agency uses Claude to produce content for you, ask how it uses it and get the answer in writing.
Update your own AI acceptable use policy
Your staff rules should point to the Anthropic Usage Policy for any Claude-based tool. A line asking staff to treat AI tools professionally costs nothing, protects the company account and matches what most firms already expect in email and chat. Our AI acceptable use policy template shows where such a rule fits.
Review customer-facing products built on Claude
If you run a chatbot or agent on Claude, your customers are “end users of products or services integrating Claude”, and the Anthropic Usage Policy applies to them. Check your own terms of use, decide how your product handles persistently abusive users, and keep logs that show the context if Anthropic ever asks about an account.
Plan for the high-risk and hardware rules
If Claude makes or shapes decisions about health, legal rights, finances, jobs or essential services, confirm that a qualified person can review and change its recommendations, and that the people affected are told AI was used. If Claude controls equipment, confirm that an operator can observe and stop it, and that it fails safe if the connection drops.
Anthropic Usage Policy: Frequently Asked Questions
When does the new Anthropic Usage Policy take effect?
On 12 November 2026. Anthropic published it on 8 October 2026. The previous version had been in force since 15 September 2025, so this is the first full refresh in more than a year.
Will the Anthropic Usage Policy get me banned for swearing at Claude?
Very unlikely. Anthropic says ordinary frustration and pushback are out of scope, and the rule targets “sustained and needless” cruelty “with no discernible purpose”. The main enforcement is Claude ending the conversation. The general clause allows warnings and account action for any breach, but Anthropic has not said it will ban users under this rule.
Does the Anthropic Usage Policy apply to API users?
Yes. It covers developers, cloud customers and the end users of products built on Claude. Anthropic has only described the end-chat feature for Claude.ai and Claude Code.
Does Anthropic claim Claude is conscious?
No. It says it is “highly uncertain” about Claude’s moral status and treats welfare measures as low-cost precautions “in case such welfare is possible”.
What else is in the Anthropic Usage Policy update?
A new section on deceptive campaigns, a narrower elections section without the blanket targeting ban, clearer weapons and surveillance rules, rewritten high-risk requirements, a new hardware safety rule and a clarified Supported Regions Policy.
References
2026 Usage Policy update (Anthropic)
Usage Policy, effective 12 November 2026 (Anthropic)
Usage Policy, effective 15 September 2025, archived version (Anthropic)
Claude Opus 4 and 4.1 can now end a rare subset of conversations (Anthropic)
Commitments on model deprecation and preservation (Anthropic)
Claude’s Constitution (Anthropic)
Anthropic bans ‘cruel’ behavior against its Claude AI (AFP via Yahoo Tech)
Anthropic bans ‘abusive or cruel behavior’ toward Claude (The Verge)
Anthropic changes usage policy to ban model abuse and election interference (TechCrunch)
Anthropic bans users from ‘needless abusive or cruel behavior’ towards Claude (The Guardian)
Anthropic asks users to stop being mean to Claude (The Register)
Anthropic bans ‘sustained and needless abusive or cruel behavior’ toward its AI models (Engadget)
Anthropic moves us one step closer to making ‘clanker’ a slur (Gizmodo)
Being mean to Claude can now get your account suspended (The Decoder)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
