Skip to content
by InfinyAI ES

IA en un minutoNewsRegulation

Regulation

Anthropic bans sustained abuse of Claude and deceiving voters in its new usage policy

IA en un minuto newsroom · Editor: Jon Elgezabal

In 30 seconds

Anthropic is refreshing the rules that limit what Claude may be used for, with the changes kicking in on November 12. The company now bars tormenting the model over and over for no reason, while ordinary complaints and experiments stay allowed. In politics, the ban now centers on manipulating voters or interfering with elections, and legitimate tailored outreach is no longer off-limits.

Anthropic has released a new version of its usage policy, the set of rules that spells out what Claude may not be used for. The policy is revised every year, and this edition comes into force on November 12. Its text says it covers every user, from people on Anthropic's own apps, such as Claude.ai or Claude Code (its coding tool), to the developers and companies that work through the API. Most of the changes, the company explains, clarify rules that already existed: they add examples of how those rules apply now that Claude does longer and more independent work, and they draw on the misuse Anthropic observed over the past year in influence operations, weapons development and surveillance.

One addition is a ban on abusive or cruel behavior toward its models when that behavior is sustained and needless. Anthropic says the rule will be applied only in extreme cases, where a person is cruel to its models again and again with no recognizable purpose. Ordinary user frustration, pushing back against the model, dark creative themes, and testing or research on models all fall outside it. Enforcement will still rest mainly on Claude's ability to end such conversations, which it can already do, in rare cases, when someone on Claude Code or Claude.ai keeps up the abuse.

The elections section gets a new name, "Do Not Undermine Democratic Processes", and now concentrates on keeping Claude from being used to deceive voters or disrupt an election. Examples include false information about candidates or about how to vote, impersonation of candidates or election officials, and attempts to drive turnout down. In parallel, Anthropic is dropping its blanket ban on personalized targeting in votes and campaigns, since that ban also blocked legitimate uses: a nonprofit writing voter information in other languages, say, or election officials notifying voters so they can cure their ballots. Reaching voters through deception or by misusing their personal data stays prohibited, now under the sections on deceptive campaigns, surveillance and privacy.

Rules against deceptive campaigns, political or commercial, used to be spread over the sections on elections, fraud, privacy and disinformation. They are now gathered in a new section called "Do Not Engage in Deceptive Campaigns or Artificial Activity". It covers concealing who is behind a message, boosting content with fake accounts or posts, and building the tools and infrastructure that influence operations need. Anthropic says it has seen commercial firms, government propaganda offices and state media outlets use Claude to set up networks of fake accounts and made-up news sites, which its policy already banned.

The policy also spells out that Claude is off limits for the software and components that make a weapon function, and for arming drones or other autonomous vehicles. Anthropic explains that it has come across attempts to use its models to write weapons guidance and control software, and that the change mirrors the way it was already applying the earlier rule. On surveillance, the policy bans tracking people without their consent, be it in real time or through analysis of data collected earlier; building or improving tools designed for surveillance; and using Claude to decide or recommend whom to investigate, arrest or charge in policing or criminal justice proceedings. The company says this does not alter what it enforces in practice either. Still allowed are tracking that people have consented to (fraud monitoring, for instance), legal research, journalism and content moderation.

Nothing changes in the requirements for high-risk uses, meaning those that can affect a person's finances, health, legal rights, livelihood or access to essential services. A qualified person must be able to review Claude's recommendations and change them if need be, and the person affected must be told that AI was used. The section has been rewritten to state which recommendations it covers and which it does not. It also adds requirements for cases where Anthropic's models are connected to machines that act on their own in the physical world and could injure someone: a qualified operator has to be able to monitor and stop them when necessary, and the machine has to be able to stay in a safe state if Claude drops out.

Anthropic has also used its supported regions page to clarify how it applies the restriction it brought in last year. Claude may not be used by people who are physically in an unsupported region, by entities incorporated or based in one, or by entities that are majority-owned or controlled by people or entities from those regions.

Why it matters · analysis and opinion

The rule that will grab the most headlines, the one on abusing the model, barely changes day-to-day use: Claude could already end those conversations, and the company itself limits it to extreme cases. What touches more people is the election part and the surveillance part. A nonprofit or an election authority that wants to write to voters in other languages gains room, while anyone who tracks people or supports police decisions now has the line drawn in more detail. For anyone connecting Claude to machines, the operator able to stop them and the safe state without a connection mean reviewing the design before the policy takes effect. And since the policy also reaches people who use third-party products built on Claude, the companies offering them have reason to go over their own terms of use.

Official source: Anthropic · Written with the help of AI: how we make the news

AI in your inbox, every day or every Friday

The stories that matter, each one in a minute. With the source for every one.

Choose one or both:

Sign up and you are in: the daily arrives every night and the weekly on Friday mornings. You can unsubscribe from any email.