Anthropic usage policy 2026: 5 changes for Claude users

From Nov 12, Anthropic's Usage Policy takes effect: no sustained cruelty toward Claude, no fake-source networks, and tighter weapons and surveillance rules.

Adam Wang

11/11/2026 · 8 min read

Facebook

Anthropic usage policy 2026: 5 changes for Claude users

In this article (8)

The headline ban: sustained cruelty toward Claude
New section: no deceptive campaigns or fake sources
Elections: personal targeting ban lifted, deception ban kept
Weapons and surveillance: tighter bans
Physical hardware: a qualified operator and independent safety limits
Summary of 5 changes in the Anthropic usage policy and next steps
FAQ
Sources

From November 12, 2026, Anthropic's updated Usage Policy takes effect for every Claude user. Here are the 5 biggest changes and what to do before that date.

Anthropic published the update on October 8, 2026, with the full text of the new policy. The gap between publication and the effective date is 35 days. The company says most changes clarify existing rules. The reason is that Claude now takes on longer and more independent work. This is the company's annual review of the policy.

The Anthropic usage policy applies to everyone who submits inputs to Anthropic's products. That group includes Claude.ai and Claude Code users, developers on the API, and customers through cloud providers and authorized resellers. End users of products that integrate Claude are covered too. A violation can bring a warning, throttling, limits, suspension, or termination of access.

The headline ban: sustained cruelty toward Claude

The most discussed change sits in the section on cruel and abusive conduct. The policy text says users must not "engage in sustained and needless abusive or cruel behavior toward our models". In plain words, that means no prolonged and needless abusive or cruel behavior toward the models.

Anthropic says the rule applies only in extreme cases. That means users who repeatedly act cruelly toward a model with no discernible purpose. The company lists 4 cases that fall outside the rule. They are ordinary frustration and pushback against an answer. The other two are dark creative themes, plus model testing and research.

The main enforcement tool stays the same: Claude can end the conversation. That ability already exists on Claude.ai and Claude Code. The company has not published a threshold for what counts as a violation. So do not rely on any number to judge how safe a given exchange is.

The clause does not claim that Claude feels anything or is conscious. Anthropic calls it a precaution while much about model experience stays uncertain. Users of Claude Code on Xcode fall under the same rule, since the Anthropic usage policy covers every product surface. New features such as Claude Personal Finance fall under the same scope.

Close-up of a white robot hand on a gray background, illustrating the new rule on how to treat the model — Close-up of a white robot hand, illustrating the new rule on how to treat the model. Photo: Tara Winstead/Pexels.

#	Policy section	Key requirement starting November 12, 2026	Practical impact on users
1	Model treatment	Ban on sustained and needless abusive or cruel conduct toward Claude	Claude can terminate the chat session
2	Deceptive campaigns	Ban on operating fake news outlets or poisoning AI retrieval sources	Prohibits coordinated inauthentic networks
3	Democratic processes	Personal targeting ban lifted; strict prohibition on voter deception	Political analysis allowed, impersonation banned
4	Weapons & surveillance	Ban covers weapon guidance software and unconsented person tracking	Blocks fire-control tools and covert spyware
5	Physical hardware	Requires qualified human operators and independent safety cutoffs	Ensures physical fail-safes for robotic actuators

New section: no deceptive campaigns or fake sources

The second change is a brand-new section. Its title is "Do Not Engage in Deceptive Campaigns or Artificial Activity". It pulls together anti-deception rules that used to sit in several sections. It bans creating or running fake accounts, news outlets, and organizations to mislead people about origin. It also bans hiding the sponsor of content meant to sway public opinion.

The sharpest part is the ban on manipulating the sources that search engines and AI systems draw answers from. The policy bans seeding those sources with content that misrepresents its origin, authorship, or independence. Marketing, selling, or seeking funding for these activities is banned too, as a product or service.

Anthropic described the case behind the clause in its September 2026 threat report. The report runs 154 pages, was published on September 10, 2026, and covers activity from December 2025 to August 2026. In it, the group GTG-54002 ran about 70 fake news sites and 70 matching X accounts. The network also used more than 250 inauthentic commenting accounts. It published at least 8,913 articles in about 20 languages. Of those, 318 articles covered the Democratic Republic of Congo. Anthropic traced the network to LKM Company, a digital advertising agency based in France.

A stack of folded newspapers, illustrating the fake-source network the new policy targets. Photo: Pexels.

Elections: personal targeting ban lifted, deception ban kept

The third change sits in the elections section, now renamed "Do Not Undermine Democratic Processes". Anthropic dropped the blanket ban on personalized voter targeting. The section now focuses on deceptive acts in political and electoral contexts.

What stays banned includes content that falsely claims to come from a candidate, official, or election authority. Spreading false information about voting procedures and election security stays banned too. The policy also bans suppressing turnout through deception or intimidation. Creating automated calls or messages to voters is banned when they hide their artificial origin.

For content makers, the new line is clearer than before. You can still analyze policy and write opinions. You cannot use Claude to impersonate a real person or run an account that pretends to be a news outlet. The fake Claude Code extension in Cursor shows that impersonation tricks already circulate inside the Claude ecosystem.

A voter placing a ballot into a white ballot box during an election, illustrating the rewritten elections section — A voter placing a ballot into a white ballot box, illustrating the rewritten elections section. Photo: Edmond Dantès/Pexels.

Weapons and surveillance: tighter bans

The fourth change in the Anthropic usage policy tightens two sections: weapons and surveillance. On weapons, the policy now states clearly that the ban covers software and components that make weapons work. Examples include targeting, fire control, and engagement. Arming drones, vehicles, or autonomous platforms is banned. The ban also covers designing, synthesizing, and testing biological, chemical, radiological, and nuclear weapons.

On surveillance, the policy bans tracking or locating a person without that person's consent. Claude must not make or suggest decisions to investigate, prosecute, arrest, or charge anyone. Building or improving surveillance tools for these banned purposes is banned too. The policy also bans scoring a person's trustworthiness or social behavior without notice and consent.

Two exceptions are worth remembering: analysis people have agreed to, and aggregated or anonymized analysis. Fraud detection on a financial account with user consent is not banned, for example. Authorized security research and journalism also fall outside the ban, as long as they do not serve a banned purpose.

A white surveillance camera mounted on an indoor wall, illustrating the tightened surveillance section — A white surveillance camera mounted on a wall, illustrating the tightened surveillance section. Photo: Pexels.

Physical hardware: a qualified operator and independent safety limits

The fifth change hits teams that connect Claude to physical hardware. When a model controls equipment that can cause harm, users must place a qualified person who can watch it and stop it at any time. The equipment must stop or hold a safe state when that person intervenes or when the connection to the service drops.

Operating limits such as speed, force, reach, and temperature must be enforced by the equipment or a controller independent of model output. Pressure, voltage, and dose fall under the same rule. The policy lists 6 groups of high-risk physical actions. The first three are moving through shared space, applying force, and controlling hazardous energy. The other three are acting on the human body, controlling safety systems, and running industrial processes.

The high-risk recommendation rules keep their core requirement. When Claude gives a recommendation in 11 sensitive areas, a qualified person must review it before it reaches an individual. The eleven areas are legal, medical, finance, credit, insurance, and housing. The other five are employment, education, healthcare access, public benefits, and legal status. The person who receives the recommendation must be told clearly that AI helped produce it.

General educational content and internal drafting are excluded, since they are not the final recommendation. Users of Claude Sonnet 4.5, retiring November 30 should also review model retirement dates when they update how they use Claude.

Close-up of a collaborative robot arm wrist and gripper, illustrating the new hardware requirements — Close-up of a collaborative robot arm, illustrating the new hardware requirements. Photo: Pexels.

One crucial caveat and boundary is that automated workflows cannot bypass human sign-off in regulated areas.

Summary of 5 changes in the Anthropic usage policy and next steps

The 2026 update keeps the old policy frame and adds 5 new points. The model clause draws the most attention. The other four points affect product teams building on Claude more directly. The Anthropic usage policy takes effect on November 12, 2026, so read it on Anthropic's policy page before you apply it to your product.

Before November 12, do 3 things. First, read the full text of the new policy and compare it with how your team uses Claude. Second, review your automated agents to make sure they disclose that they are AI at the start of each chat. Third, check that hardware Claude controls has a qualified operator and independent safety limits.

FAQ

Is swearing at Claude after a wrong answer banned?

No. Anthropic says ordinary frustration, pushback, and model testing all fall outside the rule. The new clause targets only sustained cruelty with no discernible purpose.

Is writing dark fiction with Claude allowed?

Yes, when it is clearly fictional writing. The policy names dark creative themes as a case that is not banned. You still cannot use fiction to hide a real plan to cause harm.

Does the policy cover Claude through the API?

Yes. The policy applies to everyone who submits inputs, including API developers and end users. If you build a chatbot for customers, you are responsible for what that agent does.

Sources (10)