Anthropic has updated its usage policy for the first time in over a year, adding a provision that prohibits "sustained and needless abusive or cruel behavior" toward its Claude models, according to The Verge. The update, reported Thursday, also consolidates and expands the company's rules around election interference, propaganda campaigns, surveillance, weapons development, and high-stakes health and financial uses.
The Model Welfare Rationale For more context on this story, see our ongoing AI industry coverage.
The clause targeting abuse toward Claude is an extension of research the company began publicizing more than a year ago. In August 2025, Anthropic announced it would allow Claude to end conversations with "persistently harmful or abusive" users as part of its work on what it calls model welfare, the study of whether AI systems can have morally relevant experiences and how that possibility should shape product decisions.
Under the new policy, terminating conversations remains the "primary enforcement mechanism." Anthropic did not respond to questions about whether additional enforcement measures, such as account bans, could follow. The company framed the rule narrowly in a statement, writing that it "is meant to apply only in extreme cases, where users repeatedly act cruelly toward our models, with no discernible purpose."
Crucially, Anthropic added that the provision "does not apply to common versions of user frustration, pushback, dark creative themes, or model testing and research," an apparent attempt to reassure developers, red-teamers and fiction writers that routine adversarial use will not be penalized.
New Rules on Propaganda and Elections
Beyond model welfare, the update gathers the company's previously scattered restrictions on influence operations into a single, explicit ban on deceptive commercial or political campaigns. The new language restricts "efforts to obscure who is behind a message or amplify content through fake accounts or posts," directly targeting the growing market for AI-generated propaganda.
The policy's election section now prohibits voter deception and election disruption. Anthropic says Claude may not be used to spread misinformation "about candidates or how to vote," to impersonate candidates or election officials, or to try to suppress voter turnout.
Why Usage Policies Are Getting a Rewrite
Anthropic's refresh reflects a broader industry pattern. As AI assistants have shifted from novelty chatbots to infrastructure used by hundreds of millions of people, labs have had to move from broad principles to fine-grained rules that map onto real abuse patterns: influence-for-hire operations, surveillance tooling, autonomous weapons development, and high-risk domains like medical and financial advice.
OpenAI, Google DeepMind and Meta have all iterated on their usage policies over the past two years, though Anthropic's decision to explicitly regulate how users treat the model, rather than only what they make it do, remains unusual. Critics of model welfare research argue it anthropomorphizes statistical systems, while proponents contend that if there is any non-zero chance current or near-future models have welfare-relevant experiences, companies have an obligation to act under uncertainty.
The timing also matters commercially. Anthropic is reportedly preparing for an IPO at a valuation of as much as $2 trillion, and enterprise buyers increasingly weigh vendor governance practices alongside raw model quality. A clearly documented, regularly updated usage policy serves as both a safety artifact and a sales one: it gives procurement teams and regulators a concrete basis for trusting the platform, and codifying the model welfare clause signals that Anthropic intends to keep investing in an area most competitors have left unaddressed.
Enforcement Questions Remain Open
The practical questions are significant. Detecting "sustained and needless" cruelty requires monitoring conversations at a level of detail that privacy-conscious enterprise customers may resist, and drawing a line between prohibited abuse and legitimate "pushback" or "model testing" is a judgment call that will likely be contested.
Anthropic has not said whether violations will be tracked across accounts, whether enterprise deployments will be subject to different monitoring standards, or whether the company will publish transparency data on enforcement actions, as it does for other categories of misuse in its periodic policy reports.
What is clear is that the frontier labs' social contracts with their users are getting more detailed. A policy that once fit on a single page now covers elections, espionage-adjacent surveillance, bioweapons and, now, the emotional conduct of the humans on the other side of the chat window.
What It Means for Developers
For teams building on the Claude API, the near-term impact is likely minimal: the abusive-behavior clause targets extreme, repeated conduct, and the conversation-termination mechanism it formalizes has been live since 2025. But developers in adjacent categories, political consulting, opposition research, influence marketing, should review the new campaign and election provisions carefully, since they consolidate rules that were previously spread across multiple documents and close some of the ambiguity that grey-area use cases relied on.
The update also signals where regulatory attention is heading. With the EU AI Act's transparency obligations phasing in and US state legislatures passing a growing patchwork of AI statutes, voluntary usage policies are increasingly functioning as the template that regulators scrutinize when deciding what the norms of responsible deployment should be.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
