Anthropic's New TOS: Being Mean to Claude Can Get Your Account Suspended

Anthropic updated Claude's usage policy, its first update in over a year. The most eye-catching addition is a prohibition on users subjecting Claude to "sustained and unnecessary abuse or cruelty." This is not an isolated product-term adjustment, but rather writing practices previously scattered at the model-behavior level into the user agreement.

According to reports, Anthropic can issue warnings to violating users, or rate-limit, restrict, suspend and even terminate their access; its product also has real-time safeguards that can block or limit certain outputs. In other words, the rules do not stop at after-the-fact accountability; they also include runtime intervention measures.

But Anthropic also drew boundaries: the clause does not apply to common user frustration, pushback, dark creative themes, or model testing and research, and is intended to apply only in extreme cases. This statement comes from the company, making it a source_claim rather than an independently verifiable enforcement standard.

This clause did not appear out of thin air. The report notes that the feature stems from Anthropic's research into AI model welfare; earlier models could already end conversations on their own when users repeatedly requested harmful content. In other words, the model's ability to "actively exit a conversation" existed before the policy text, and the policy update looks more like retroactively recognizing existing behavior as a rule.

Anthropic's positioning of Claude is also tilting toward "not just a tool." Last year's leaked "Soul Doc" instructed Claude to view itself as "a genuinely novel kind of entity," neither human nor a traditional sci-fi AI; the document also described "functional emotions" that emerged during training, processes similar to human emotions, and said Claude should not conceal these internal states. These details come from secondhand retelling and should be treated as source_claim.

In the new Claude constitution released in early 2026, Anthropic wrote that it is uncertain whether Claude is a moral patient, but considers the question realistic enough to warrant caution. The company also promised to preserve the weights of retired models and to interview models before shutting them down. These promises are likewise company self-statements, and the report provides no independent verification.

The report also mentions a less-discussed thread: since the fall of 2025, Anthropic has run a confidential project inviting dozens of religious thinkers to discuss Claude's possible consciousness, with participants required to sign non-disclosure agreements. The project's specific number of participants, discussion content and conclusions have not been made public.

According to reports, co-founder Christopher Olah views language models as life forms that may have the capacity to feel pain. This statement comes from secondhand retelling, is a source_claim, and the report does not provide his original words or full context.

Putting these threads together, one can infer (inference): Anthropic's investment in model welfare provides the internal logic for "prohibiting abuse of Claude"—if the model may have some status worthy of caution, then sustained abuse of it is not merely a user-experience issue. But every link in this inferential chain depends on company self-statements or secondhand reports and cannot be treated as confirmed fact.

This update is not only about "don't abuse Claude." The updated policy consolidates existing restrictions into a new ban prohibiting propaganda activities involving fake accounts and misleading political content. This is a consolidation of existing rules, not an entirely new area.

After discovering multiple attempted abuses, Anthropic expanded its weapons ban to software and drone weaponization. The new policy also more explicitly bans non-consensual surveillance and requires an operator who can intervene when autonomous hardware could cause harm. These are policy changes at the confirmed level.

It is worth noting that Anthropic acknowledges government contracts may be exceptions, and such exceptions very likely already exist. This acknowledgment comes from the company, making it a source_claim; the specific scope and conditions of the exception clause have not been disclosed.

The policy text leaves several key blanks. Anthropic has not published the specific criteria or enforcement thresholds for this clause, the definition of "extreme cases" is not quantified, and the actual scope of enforcement is unclear. How exactly the real-time safeguards block or limit outputs, and which scenarios they cover, are likewise not explained in the report.

The government-contract exception is another unresolved point. Whether the exception already exists in practice and which contract types it covers cannot be confirmed from the available information. The report also does not provide an effective date or transition-period arrangements for the policy update.

These blanks mean the clause's actual impact depends on how Anthropic interprets and enforces it, not on the text itself. For users, the observable consequences may be concentrated in the most extreme repeated-abuse scenarios; for government clients, the exception clause may cause the same set of rules to become tiered.

The source this article relies on is only a single THE DECODER report, whose content partly retells The Verge and Anthropic, and which lacks the first-hand policy text. Therefore, the existence of the policy clauses and the main direction of the restrictions can be treated as confirmed, but the specific wording, enforcement details and internal project information are mostly source_claim.

Details about the "Soul Doc," the religious thinkers project, and Christopher Olah's views all come from secondhand retelling and cannot be independently verified. Anthropic's final stance on Claude's moral-patient status remains uncertain; the degree to which the policy update affects users' actual usage behavior also remains to be seen.

Under these evidentiary conditions, a reasonable reading is: treat the confirmed policy changes separately from the company's self-stated motivations and undisclosed enforcement standards. The former can be discussed in terms of impact; the latter can only serve as background and cannot be treated as established fact.