Anthropic changes usage policy to ban model abuse and election interference

Anthropic published its annual usage policy update on Thursday, with the new rules taking effect November 12. Most of it clarifies existing rules for a Claude that now does longer, more independent work. The company says the changes draw on misuse patterns it has seen in influence operations, weapons development and surveillance, and add controls for when Claude takes physical actions on its own.

Two changes stand out. First, an elections section renamed Do Not Undermine Democratic Processes, focused on deceiving voters, impersonating candidates or officials, and suppressing turnout. Second, a new prohibition on sustained and needless abusive or cruel behavior toward the models themselves, which Anthropic says applies only in extreme cases with no discernible purpose.

Start with elections, because that is where the stakes for the public are highest. We think the new section is better policy than what it replaces. It names concrete harms: automated messages that hide their artificial origin, fake accounts or outlets posing as real campaigns or news organizations, false information about how to vote, and efforts to intimidate people out of participating. Those are the abuses that actually threaten an election, and naming them precisely is more enforceable than vague political-content bans.

Anthropic also dropped a blanket ban on personalized voter and campaign targeting, saying it had caught legitimate civic work like nonprofits writing voter information in other languages or officials sending ballot cure notices. We understand the reasoning and broadly agree. But be clear about the trade: loosening targeting rules ahead of an election means more political actors will use Claude for outreach, and the line between persuasive targeting and deceptive targeting will be drawn case by case, by the company, in real time. I'm watching how that enforcement holds up when the pressure arrives.

Now the part everyone will talk about. Claude has been able to end persistently abusive conversations since an earlier update. The new policy goes further and bars users from pursuing those conversations. Anthropic is careful that this does not cover frustration, pushback, dark creative themes, or testing and research. In practice, ending the chat stays the main enforcement tool, according to reporting on the update.

We are of two minds. There is a decent argument that tolerating purposeless cruelty toward a humanlike system is corrosive for the people doing it, and that a company can set norms for its own product. The more contested argument sits underneath: TechCrunch notes the change follows Anthropic's outreach to religious scholars, in some cases trying to convince them Claude could be conscious or have a soul. Writing model welfare into terms of service moves a deeply unsettled question from philosophy seminars into account enforcement. That deserves scrutiny, not mockery and not deference.

Our read: the election and surveillance rules are the substance here, and they are mostly an improvement. The abuse clause is narrow on paper. Where this leads, if other labs follow, is a world where user conduct toward AI is policed on the AI's behalf. That should be argued in the open before it becomes normal.

Anthropic says the update reflects its latest threat intelligence and clarifies requirements for high-risk uses in areas like health and finance. Separate sections of the policy also address surveillance and law enforcement uses, including bars on using Claude to make decisions about investigating, arresting, charging or sentencing people.