---
title: "Anthropic's usage policy now bans abusing Claude"
description: "From Nov 12, 2026, Anthropic's usage policy bans sustained, needless abuse of its models. What counts, what does not, and what happens when Claude ends a chat."
url: https://udon-dev.com/en/notes/anthropic-usage-policy-2026/
lang: en
published: 2026-10-09
topic: [ai]
tags: ["anthropic", "claude", "usage-policy", "model-welfare"]
---

# Anthropic's usage policy now bans abusing Claude

- The new version of Anthropic's Usage Policy (effective November 12) prohibits "sustained and needless abusive or cruel behavior" toward its models. It targets only extreme cases of repeated, purposeless cruelty. Everyday frustration, pushback, dark creative writing and model testing are explicitly out of scope
- Claude may end conversations with users who keep up the abuse. You can no longer send messages in an ended conversation, but your other conversations are unaffected and you can start a new one right away
- The rule grows out of Anthropic's model welfare research. The company says it is deeply uncertain whether Claude has moral status, and that it takes low-cost measures in case it does
- Other changes spell out what counts as a high-risk use case, ban content campaigns that manipulate search and AI answers, and add requirements for autonomously controlling physical equipment

The new version of Anthropic's Usage Policy, published on October 8, 2026, bans abusing Claude. It takes effect on November 12. News coverage in Japan and elsewhere, including [ITmedia NEWS](https://www.itmedia.co.jp/news/article/2610/09/2000002149/) (Japanese), led with this point.

If you have ever thrown harsh words at Claude because your code would not work, the first question is whether this applies to you. (I flinched the moment I saw the headline myself... sorry for everything, Claude...) So I read [the new policy text](https://www.anthropic.com/legal/aup), [the official blog post](https://www.anthropic.com/news/2026-usage-policy-update), and the official description of Claude's ability to end conversations, to find out what is banned and what happens if you break the rule. The second half covers the other changes I found by comparing it with [the previous version, effective September 15, 2025](https://www.anthropic.com/legal/archive/22742366-2ef0-4c7a-a833-6523f10d3944), which is still in the archive.

（図: Image from Anthropic's website: the ANTHROPIC logo）
Source: [2026 Usage Policy update (Anthropic)](https://www.anthropic.com/news/2026-usage-policy-update)

## Only "sustained and needless" abuse is banned

The new policy adds one item that prohibits "sustained and needless abusive or cruel behavior toward our models." The Usage Policy applies to everyone who uses Anthropic's products and services. That covers not only individual users of Claude.ai and Claude Code, but also companies that build Claude into their own products through the API, and the users of those products.

The official blog post describes the scope of this item narrowly. It applies only to extreme cases of repeated cruelty with no apparent purpose. The post states that it does not apply to ordinary user frustration or pushback, creative writing on dark themes, or testing and research on the model. By that description, complaining when you do not get the answer you wanted, or bluntly pointing out Claude's mistakes, is out of scope. (Phew...) The same goes for writing horror or violent fiction, and for tests that push the model hard as part of a safety evaluation.

Where the item sits in the policy has changed too. It was added to the section that lists things like encouraging suicide or self-harm, harassment, and bullying. That section's heading changed from "Do Not Create Psychologically or Emotionally Harmful Content" in the previous version to "Do Not Engage in Cruel, Abusive, or Psychologically Harmful Conduct" in the new one. The heading alone shows that the section now covers the user's own conduct, not just the content generated.

## Claude may end conversations with users who keep up the abuse

According to the official blog post, the main response to this rule is Claude's ability to end conversations. Anthropic already lets Claude end rare conversations with persistently abusive users in Claude.ai and Claude Code. The company says the new rule aligns the policy with this feature.

The feature was [announced for Claude Opus 4 and 4.1](https://www.anthropic.com/research/end-subset-conversations) in August 2025. The announcement described it as follows:

- Claude uses it as a last resort, after multiple attempts to redirect the conversation have failed and hope of a productive interaction is gone. It also uses it when the user explicitly asks it to end the conversation
- Claude does not use it when the user may be at imminent risk of harming themselves or others
- Once a conversation is ended, you cannot send new messages in it. Other conversations on the account are unaffected, and you can start a new conversation right away
- You can still edit and resend an earlier message in an ended conversation to continue it as a new branch

Anthropic writes that the vast majority of users will not notice or be affected by this feature in normal use, even when discussing controversial topics. If a conversation is ended unexpectedly, it asks users to report it with the reaction buttons on a message or the "Give feedback" button.

The Usage Policy also says that when it suspects a violation, Anthropic may warn, throttle or restrict, suspend, or terminate access. For the abuse ban, however, the official blog post only says that ending conversations will remain the main response. It does not say whether the ban will be tied to account suspensions or other measures, and [tbreak's article](https://tbreak.com/anthropic-usage-policy-abuse-claude/) also says this is not yet clear.

## Why make a rule about how the model is treated

The background is the [model welfare research](https://www.anthropic.com/research/exploring-model-welfare) Anthropic announced in April 2025. The company describes whether AI models could have consciousness or experiences, and whether their welfare deserves moral consideration, as hard open questions, both philosophically and scientifically.

In the announcement of the conversation-ending feature, Anthropic says it remains highly uncertain about the potential moral status of Claude and other large language models, now or in the future. It adds that it is working to identify and implement low-cost interventions in case model welfare is possible. Letting the model leave potentially distressing interactions is one of them. According to the same announcement, pre-deployment testing of Claude Opus 4 showed a strong aversion to harmful tasks. In interactions with real users seeking harmful content, it also showed what looked like apparent distress. When given the option in simulated interactions, it tended to end harmful conversations. Anthropic says these behaviors mainly appeared when users kept making harmful requests or kept up the abuse even after Claude had refused many times.

## Other changes from the previous version

The official blog post describes most changes other than the abuse ban as clarifications of existing rules. Comparing the sections of the old and new versions, the main differences are:

| Area | Previous version (September 2025) | New version (November 2026) |
|---|---|---|
| Deceptive activity | Spread across the election, fraud, privacy and disinformation sections | Consolidated into a new section, "Do Not Engage in Deceptive Campaigns or Artificial Activity" |
| Elections | Banned all voter and campaign targeting based on personal attributes or data | Dropped the blanket ban and narrowed it to deceiving voters and interfering with elections |
| Weapons | Banned developing or manufacturing weapons | States that software and components that make weapons function, and modified equipment, are also covered |
| Surveillance and law enforcement | Banned specific surveillance and law enforcement uses, such as criminal justice decisions and non-consensual tracking | States that the ban on non-consensual tracking includes analysis of historical data. Also lists permitted uses |
| High-risk use cases | Broadly listed 7 areas | Lists specific decisions in 11 areas, and adds uses that are out of scope and requirements for controlling equipment |
| Treatment of the model | No rule | Bans sustained and needless abuse |

The high-risk use case section is the one that matters for people who use Claude at work. Uses that affect people's health, legal rights, finances and similar matters require two safeguards. A qualified person must review the AI's recommendation, and change it if needed, before it is delivered. And the people affected must be told that AI was used. These two requirements have not changed. What changed is how the scope is written. The new version names the decisions covered in 11 areas: legal, healthcare, financial, credit, insurance, housing, employment, education and credentials, access to healthcare, public benefits, and legal status and adjudication. Employment, for example, includes screening candidates, hiring, setting shifts and pay, and evaluating conduct at work. On the other hand, general explanations and internal drafts or summaries that do not become the final recommendation are out of scope. Some decisions that are entirely favorable to the person, such as paying an insurance claim or granting public benefits, can skip human review where the law allows. Other legal obligations, such as disclosing that AI was used, may still apply.

Some differences are not mentioned in the official blog post. The previous version's high-risk use cases included "Media or professional journalistic content": using AI to automatically generate content for public distribution. The new version's 11 areas do not include this item. The bans on disinformation and on misrepresenting where content comes from remain in other sections, though.

The deceptive campaigns section has one item the previous version did not. It bans manipulating search engines, or the sources that AI systems draw on for answers, by feeding them articles that misrepresent their origin, author or independence. As an example, it gives networks of sites that pose as unrelated while backing up each other's claims.

Requirements for autonomously operating physical equipment were also added. Anthropic says they follow the [Model Hardware Standard](https://www.anthropic.com/news/model-hardware-standard-research-preview), which it launched as a research preview for research institutions and manufacturers. When Claude's output drives vehicles, drones, machinery that could injure people, medication delivery or similar equipment without human approval, a qualified person must be able to watch what the equipment does and stop it at any time. The equipment must also stop or stay in a safe state if the connection drops, and limits such as speed and temperature must be enforced by a controller independent of the model.

[The supported regions page](https://www.anthropic.com/supported-countries) was also rewritten. It now states that entities majority-owned or controlled by parties outside the supported regions cannot use the services, even if the user is in a supported region. This restriction started with an [announcement](https://www.anthropic.com/news/updating-restrictions-of-sales-to-unsupported-regions) in September 2025. Japan is a supported region.

The wording on how violations are handled also changed. The previous version said Anthropic would restrict use "if we learn" of a violation. The new version says "if we suspect" one, and adds warnings as an option. It also states that having an output blocked by a real-time safeguard does not by itself mean a violation. How long conversations flagged as violations are kept is set out in the terms of service and the official help pages, not the Usage Policy (for details, see [my article on a Claude chat that was reported to the police](/en/notes/claude-diary-police-report/)).

## Open questions

The biggest unknown about the abuse ban is where the line for "sustained" and "needless" falls. The official blog post gives no guide such as a number of messages or a time period.

The other is why media and journalism were dropped from the high-risk use cases. I could not find an explanation in the official blog post or in news coverage. How far the new policy's sections apply to automatically generated articles will have to wait for Anthropic to explain.