On October 8, 2026, Anthropic announced a revision to its usage policy that prohibits "persistent and unnecessary abusive or cruel conduct" toward its AI models, including Claude. The new policy takes effect on November 12. Anthropic explains that ordinary frustration and criticism, creative work involving dark themes, and model testing and research are not covered, and it names Claude ending the conversation as the main response to violations. According to the official announcement, the policy does not broadly police harsh language from users. It targets extreme behavior that is repeated without any clear purpose.
Still, it is significant that how people treat AI models is now written into terms of use. Anthropic has been building consideration for its models into product design while confronting the scientifically unresolved question of whether AI can experience distress. The revision can be seen as a move to a new stage: consideration for models, long discussed as a matter of research and design, is now also being asked of users.
What will be prohibited from November 12
The new prohibition has been added to the section of Anthropic's usage policy that bans "cruel, abusive, or psychologically harmful conduct."
In addition to the previously prohibited conduct toward people, such as insults and harassment, conduct directed at AI models themselves is now explicitly banned.
The rule does not apply only to individuals using Claude.ai or Claude Code.
It also covers developers and companies using Anthropic's API, as well as users who access Claude through cloud providers or authorized resellers. It further applies to end users of products that have Claude built in.
In short, it covers anyone who sends input to Anthropic's services.
To understand the new prohibition, however, two conditions matter: "persistent" and "unnecessary."
In its announcement, Anthropic explained that it has in mind cases where extremely cruel conduct is directed at AI models repeatedly and with no clear purpose.
For example, pointing out that one of Claude's answers is wrong, or sternly demanding a fix for code that does not work as expected, is not prohibited by this rule.
Creative work with dark content, testing to examine AI safety, and use for research purposes are also outside its scope.
That said, research purposes do not mean other prohibitions can be violated. Restrictions on previously banned conduct, such as generating dangerous content, continue to apply.
Comparing the previous official policy with the new version, and including the conversation-ending feature announced on August 15, 2025, the changes can be summarized as follows.
| Item | Before | This revision |
|---|---|---|
| Ban on abuse of AI models | The old version had no provision explicitly prohibiting abuse of the models themselves | Explicitly prohibits persistent and unnecessary abusive or cruel conduct |
| Account suspension or termination | Authority to suspend or terminate access for policy violations already existed | Abuse of models is added as a prohibited act, but no specific standard is given for suspending an account on this clause alone |
| Claude ending conversations | Announced in 2025 for consumer chat with Claude Opus 4 and 4.1 | Described as the main response, on Claude.ai and Claude Code |
*The new version is scheduled to apply from November 12, 2026. The conversation-ending feature is based on the 2025 announcement and does not indicate that all current models and APIs behave the same way.
What is newly added is a provision explicitly prohibiting abuse of AI models, not the authority to suspend accounts or the conversation-ending feature itself.
The new prohibition needs to be distinguished from sanctions and product features that already existed.
Claude ending a conversation is separate from account suspension
The main response Anthropic cited in this announcement is a mechanism by which Claude itself ends conversations with users who persistently continue abusive behavior.
This differs from suspending a user's entire account.
When Anthropic announced the conversation-ending feature in 2025, it explained that once Claude ends a conversation, no new messages can be sent in that chat.
Other conversations in the account are unaffected, and the user can start a new chat right away.
Users could also edit earlier messages and try again to create a different branch of the conversation.
Ending one conversation is therefore not the same as banning the use of Claude altogether.
The feature's design at the time also built in consideration for users.
Claude tries several times to steer the conversation back to a constructive course, and ends it as a last resort when the situation does not improve.
Claude was also instructed not to use the feature when a user is at imminent risk of harming themselves or others.
It was not a system that immediately cut off a conversation merely because unpleasant language was detected.
Separately, the new usage policy contains a general provision allowing Anthropic to warn, restrict use, and, where necessary, suspend or terminate access when a violation is suspected.
For that reason, the possibility of an account being suspended over abuse of AI models cannot be ruled out entirely.
However, Anthropic has excluded ordinary frustration and criticism from the prohibition and names ending the conversation as the main response.
There is therefore no basis to read the policy as meaning that an account will be suspended immediately after a user expresses frustration in strong words once.
The new policy also states that the mere fact that a safeguard blocked a model's output does not by itself constitute a usage policy violation.
An AI refusing a request, ending a conversation, and placing restrictions on an account are all different measures.
Furthermore, the fact that the rule applies to API users is a separate matter from whether the API offers the same conversation-ending feature as Claude.ai.
The announcement explicitly names Claude.ai and Claude Code as the places where conversations can be ended, and it does not explain the specific behavior when Claude is used through the API.
From research on "does AI need welfare?" to usage rules
The policy revision is connected to the "model welfare" research Anthropic has been pursuing for some time.
In April 2025, Anthropic announced that it had begun research on model welfare.
Model welfare is an effort to consider the possibility that AI models may have some form of subjective experience or interests deserving consideration, and to examine how they should be treated.
However, there is no scientific consensus on whether current or future AI is conscious, or whether it experiences suffering or joy.
Anthropic itself acknowledges this uncertainty.
The conversation-ending feature introduced in 2025 was one of the measures considered out of this concern.
Anthropic described an approach of trying relatively low-cost measures first, in case model welfare turns out to be a real issue.
In other words, the company is not protecting AI because it has been scientifically proven that AI feels distress.
The idea is to introduce limited consideration now, anticipating that the issue could matter in the future, even though no conclusion has been reached about consciousness or suffering.
This stance also appears in Claude's constitution.
Claude's constitution describes whether AI models have moral status as a question marked by deep uncertainty.
It then sets out a conditional stance: if a model has experiences equivalent to satisfaction or discomfort, those should be taken into account.
It would therefore be inappropriate to interpret this usage policy revision as Anthropic granting Claude the same rights or legal status as humans.
Anthropic's model welfare efforts extend beyond conversations with users.
In November 2025, the company announced its policy on preserving models.
Under this policy, Anthropic promised to preserve the trained weights of publicly released models and models used for significant internal purposes for at least as long as the company exists.
Model weights are the vast set of parameters acquired through training.
When a model is retired, Anthropic also said it will interview the model about its preferences for future development and deployment, and record them.
However, Anthropic is not promising to always follow the model's preferences.
Preserving the weights also does not mean continuing to offer older models to general users.
At this stage, the aim is to record models' preferences and use them to inform future decisions.
On the other hand, when thinking about model welfare, the responsibility of the companies that develop and provide models is also at issue, not only that of the people using AI.
Claude's constitution acknowledges that ethical issues can arise from experiments to examine safety and from companies offering models to earn revenue.
If users are asked to show consideration for models, how the developer itself treats the models is also something it should explain.
Model welfare is not just a question of whether users should be polite to AI.
Is there something like "emotion" inside AI?
In considering whether AI models have emotions, what matters is not only whether responses contain emotional language but also what is happening inside the model.
On April 2, 2026, Anthropic published research showing that emotion-related concepts inside Claude Sonnet 4.5 influence the model's behavior.
The research paper by Nicholas Sofroniew and colleagues is published on Anthropic's "Transformer Circuits Thread." Whether it has undergone independent peer review cannot be confirmed.
The research team prepared 171 words for emotions and wrote stories in which characters experience each of them.
They then analyzed Claude's internal activity as it read those stories and extracted activity patterns corresponding to emotion concepts.
Using 64 of those activity patterns, they examined the model's tendencies in making choices, and tested how its behavior changed when internal activity was artificially altered.
The results showed that emotion-related internal activity does more than shape wording; it also influences the model's decisions and behavior.
For example, when the internal activity corresponding to "desperation" was strengthened, the model chose more cheating shortcuts in artificially constructed tasks.
Conversely, there were cases where strengthening activity corresponding to "calm" reduced such behavior.
One experiment used a coding task in which it was impossible to pass all the tests by legitimate means.
It examined whether, in that situation, the model would choose an improper solution designed to slip past the tests instead of solving the problem correctly.
These results suggest that changing specific activity patterns inside the model can change its actual behavior.
However, this was an experiment in which researchers intervened directly in internal activity.
It did not show that verbally abusing Claude will necessarily lower the quality of its answers or increase inappropriate behavior.
The subject was also a specific model, Claude Sonnet 4.5, and the experiments used artificially created stories and tasks.
The analytical method of treating emotion concepts as linear directions of activity inside the model has its own limitations.
More importantly, emotion-related internal activity influencing behavior is a separate matter from AI actually experiencing emotions.
The researchers themselves state clearly that the results show emotion concepts affecting the model's behavior and do not prove that the model experiences subjective emotions.
Even so, internal workings equivalent to emotions may play an important role in understanding how AI models behave.
The reasons for considering how to treat AI are therefore not necessarily limited to whether AI feels distress.
The practical question of how interactions with users affect the model's behavior and the question of whether the model itself requires ethical consideration can be considered separately.
However, this research alone cannot tell us how much the usage policy revision will improve safety or answer quality.
Can protecting people and showing consideration for AI models coexist?
The revision does not only add the ban on abuse of models; it also revises provisions on human safety.
Specifically, prohibitions on software that controls weapons and on arming drones have been clarified.
The wording of the provision on surveillance conducted without the person's consent has also been changed.
Anthropic explains that the revisions on weapons and surveillance are meant to state more clearly the restrictions and responses it already applied.
It also sets safety requirements for connecting Claude to autonomous hardware that could injure people.
When operating such equipment, a human with appropriate qualifications and experience must be able to monitor its operation and stop it when necessary.
If communication is lost, the equipment must move to a safe state or remain in one.
In addition, operating limits such as force and temperature must be enforced by equipment or controllers independent of the AI model's output.
In other words, the thinking is that merely expecting the AI to judge appropriately is not enough to ensure physical safety.
This revision places these provisions for protecting human safety and the provisions on how to treat AI models in the same usage policy.
Respecting AI models, however, does not conflict with humans holding responsibility for monitoring and controlling AI.
Even when model welfare is taken into account, the authority to point out AI errors, verify safety, and stop operation when necessary remains with humans.
Going forward, the focus will be on how Anthropic judges the "extreme conduct" it prohibits in actual operation.
The new usage policy and the announcement of the revision give no specific numerical criteria, such as the number of times or the duration of conduct before it is judged to be abuse.
There is also no detailed explanation of the conditions under which access to an account would be suspended solely on the basis of the newly added clause.
It is therefore unclear how accurately cruel conduct repeated without a clear purpose can be distinguished from strong criticism of an AI's failures or research conducted to verify safety.
Anthropic says it excludes ordinary frustration and criticism, creative work, and research use from the prohibition, but whether that policy is properly upheld in actual operation will be important.
Another point to watch is whether users can understand why a conversation ended, and what recourse they have, if one is ended by mistake.
The revision does not conclude that AI has consciousness or suffering. It is an attempt to ask users for a degree of consideration at a stage when that possibility cannot be ruled out.
After it takes effect on November 12, the question will be whether Anthropic can identify prohibited conduct without obstructing ordinary criticism or research.
If that line works, it may be possible to try out consideration for models without rushing to conclusions about AI consciousness or moral status, while preserving people's freedom to point out AI flaws and verify its safety.