The rule that drew the most attention is the one with the least money attached. Anthropic, in a rewrite of its usage policy published Wednesday, will bar users from subjecting its Claude chatbot to what it calls sustained and needless abusive or cruel behavior.
The provision is narrower than it sounds. It applies only to extreme cases, the company said, where a user repeatedly acts cruelly toward the model with no discernible purpose. It does not cover ordinary frustration, pushback, dark creative themes, or model testing and research.
The clause formalizes a step Anthropic took in August, when it allowed Claude to end conversations with persistently harmful or abusive users as part of research into what the company calls model welfare. Ending the conversation will remain the primary enforcement mechanism. Anthropic would not say whether it would go further, such as banning users.
The cruelty rule sits inside a broader update – the first in more than a year – that tightens restrictions on election interference, weapons and surveillance. The company expanded its weapons ban to cover the software and components that make weapons work, including systems for arming drones and other autonomous vehicles.
Anthropic also added new limits on high-risk uses in health care and finance, and folded several scattered rules into a single ban on deceptive campaigns. That section targets efforts to hide who is behind a message, or to amplify content through fake accounts and fabricated news outlets.
The elections section was renamed to something plainer: Do Not Undermine Democratic Processes. It is now focused on deceiving voters – spreading false information about candidates or how to vote, impersonating candidates or election officials, and suppressing turnout. In a reversal, Anthropic dropped its blanket prohibition on personalized voter targeting, while keeping rules against deceptive targeting and the misuse of personal data.
The changes take effect November 12.
The weapons and surveillance changes are the most commercially significant. The company said it has seen multiple attempts to use Claude to develop guidance for and control software for weapons, which prompted the expansion. The surveillance rules bar using the model to track individuals without consent, to build surveillance tools, or to recommend who should be investigated, arrested or charged.
The deceptive-campaigns section consolidates rules that were previously scattered. It covers efforts to obscure who is behind a message, to amplify content through fake accounts or posts, and to build the tools and infrastructure for running influence campaigns, whether political or commercial.
On the same day, Anthropic launched a cybersecurity effort aimed at critical infrastructure. A program called the Critical Infrastructure Defense Program will offer operators of power grids, water systems, transportation networks and government agencies access to its frontier models, embedded engineers and threat research. A separate free service, OSS Scans, will scan open-source software for vulnerabilities.
The two announcements share one calculation. As its models grow more capable, Anthropic is under pressure to show regulators and enterprise customers that it can police how they are used, and to do so before governments write the rules for it.
The cruelty clause is the unusual piece. Companies routinely ban harmful uses of their products. Almost none write rules about how a user may treat the product itself. Anthropic frames it as a last-resort measure aimed at a real but small phenomenon: people who torture a chatbot for no reason.
The company is careful about the wording. The line between a policy against abuse and a claim that the model has feelings is one Anthropic does not cross. Its executives describe the rule as a matter of responsible deployment, not of attributing any kind of inner life to Claude.
The change has drawn a mix of amusement and criticism online, where users of rival chatbots have long treated goading a model into distress as a minor sport. Anthropic’s answer is that the rule is aimed at behavior far beyond taunting – repeated, purposeless cruelty – and that ordinary use will be unaffected.
The election and surveillance changes matter more to the company’s commercial future. Governments and large buyers increasingly demand written assurance that a model will not be used to sway voters or track citizens. A policy that names those limits explicitly is, in part, a sales document.
Anthropic is also trying to stay ahead of a regulatory curve. Lawmakers in Europe and the United States are writing rules for artificial intelligence, and companies that publish their own stricter limits are often given more room to keep operating while those rules take shape.
The timing is unlikely to be accidental. The update arrives as AI companies face growing scrutiny over their role in political campaigns and their willingness to sell to military and intelligence customers. By drawing its own lines first, Anthropic is staking out a position before regulators draw them.
None of this changes what Claude can actually do. The model’s abilities are unchanged; only the rules around its use have shifted. What changed is the document, and in a young industry where trust is still being earned, the document is often the product.
