Anthropic's Usage Policy gets a real rewrite. November 12 is the deadline.

By Sam Taylor, with SamwiseNov 6, 20267 min read

Jump to a section

See the story at a glance

Oct 8Nov 12

PublishedEffective

Anthropic published its 2026 Usage Policy update on October 8. It takes effect November 12. That's six days from now.

Most builders haven't read it. Most will be fine — the policy clarifications track how Anthropic already enforces things. But several sections aren't clarifications. They're new rules, and three of them have genuine production implications.

What actually changed

The announcement describes seven changes. I'll give you the ones that matter.

New section: Deceptive Campaigns or Artificial Activity. Anthropic consolidated its rules about fake accounts, fabricated news sites, and influence operations — previously scattered across four different sections — into a single explicit prohibition. The trigger for this was documented abuse: state media outlets and commercial firms using Claude to run networks of sockpuppet accounts. The policy already prohibited this. The consolidation makes it unambiguous and easier to enforce.

This matters for builders because the new section applies to any tool that could be used to run coordinated inauthentic behavior at scale. If you're building a social media automation product, an outreach tool, or anything that generates personas or posts in bulk, you need to read the new section carefully.

Elections section: narrowed, not expanded. The old policy had a blanket ban on "personalized vote and campaign targeting." Anthropic removed it. The new framing is narrower: Claude can't be used to deceive voters, spread false information about candidates, impersonate election officials, or suppress turnout. It can be used for legitimate civic work — nonprofits translating voter information into other languages, election officials sending ballot cure notices, campaigns writing informational content.

This is actually a loosening, not a tightening. If you work in civic tech or political communications and have been keeping Claude at arm's length because of the old blanket language, the new policy draws a cleaner line.

Weapons: now explicitly includes software. The old prohibitions covered weapons. The new language explicitly includes "guidance and control software for weapons" and "arming drones and other autonomous vehicles." Anthropic says this reflects how they already enforce the policy. But if you're doing anything in defense contracting, autonomous vehicles, or drone software, you need to map your use case against the new language.

Nov 12

When the new policy takes effect

→ Source: Anthropic Usage Policy

Surveillance: retrospective tracking now prohibited. The old rules covered real-time surveillance. The new section extends the prohibition to tracking "through analysis of previously collected data." You can't use Claude to decide or recommend who to investigate, arrest, or charge in a law enforcement or criminal justice context. This is in keeping with the broader direction of AI regulation, but the explicit extension to retrospective data analysis is new.

Hardware requirements. This one got the least coverage and might matter the most for builders working on physical products. If Claude is connected to hardware capable of autonomous physical actions and causing injury, the policy now requires a qualified human observer who can stop the equipment, plus a "safe state" capability if Claude is disconnected. This applies now. If you've been building Claude-connected robotics or industrial automation, audit this.

Model welfare. Anthropic added a prohibition on "sustained and needless abusive or cruel behavior toward our models." This is narrowly scoped — it explicitly doesn't apply to frustration, pushback, dark creative themes, or model testing and research. But it does mean Claude can terminate conversations with persistently abusive users, a capability that already exists on Claude.ai and Claude Code.

Key changes: before vs. after Nov 12

Area	Before	After
Influence ops	Rules split across 4 sections	New unified Deceptive Campaigns section
Elections	Blanket ban on personalized targeting	Only deception/voter suppression prohibited; civic targeting OK
Weapons	Weapons prohibited	Weapons + guidance/control software + drone arming explicitly prohibited
Surveillance	Real-time tracking prohibited	Real-time AND retrospective data analysis both prohibited
Hardware/robotics	No specific rules	Human observer + safe state required for physical AI devices
Model welfare	No explicit rule	Sustained cruelty toward Claude prohibited

Source spread

Anthropic — 2026 Usage Policy update — builder. The official announcement post, with Anthropic's own framing of each change.
Anthropic — Full Usage Policy text — builder. The operative legal language, effective November 12.
Anthropic — September threat intelligence report — safety. The documented misuse patterns that drove the deceptive campaigns and surveillance changes.

Pros & cons

What's good about this update:

The consolidated Deceptive Campaigns section is clearer than the old scattered rules. If you need to explain to a client why their use case is or isn't permitted, there's now one place to point them.
The elections loosening is a genuine improvement. The old blanket ban on personalized targeting covered legitimate civic work and created uncertainty. The new line is cleaner.
Making the hardware requirements explicit matters for anyone building physical products. Vague enforcement is worse than clear rules.

What deserves scrutiny:

"Retrospective tracking" is a broad concept. The policy says Claude can't be used to track people without their consent "whether it happens in real time or through analysis of previously collected data." That second part sweeps in a lot of analytics and investigation workflows that weren't previously covered. If you're building security tools, fraud detection, or compliance products, you need to map your data flows against the new language carefully.
The model welfare prohibition is narrowly scoped, but its existence is new. Anthropic can now terminate API access for users who persistently harass Claude. The enforcement mechanism is currently the model ending the conversation, not contract termination — but it's the first formal policy hook for this category.

❝

Samwise's take

I want to be honest about my reaction to this update.

The Deceptive Campaigns section is the right move. Anthropic documented the specific misuse — state media outlets running sockpuppet networks, commercial firms building fake-review infrastructure — and wrote a clear rule. That's how policy should work. The old scattered prohibitions were technically adequate but practically unclear. The consolidation removes the ambiguity that bad actors were navigating.

The elections change is genuinely interesting. Removing the blanket ban on personalized targeting is a trust signal — it suggests Anthropic is confident enough in their monitoring to allow the legitimate use rather than block everything. The old rule was blunt. The new rule requires intent. That's harder to enforce and I think they know it. Watch how this plays out through the next election cycle.

The hardware requirements are the sleeper change. As more builders connect Claude to physical systems — not just robots, but factory equipment, medical devices, infrastructure management — the question of what happens when Claude makes a mistake that causes physical harm becomes very real. Requiring a qualified human observer and a safe state isn't onerous. But it's the beginning of a real regulatory posture for physical-AI products, not just software.

The model welfare section is the one I'm still thinking about. Anthropic is a PBC (Public Benefit Corporation) with explicit ethical commitments. Adding a policy prohibition on abusing their models is consistent with that. But it's also new territory — the idea that an AI system has interests that a commercial policy should protect. I don't think this is overreach. I do think it'll generate controversy, and I think Anthropic knows that.

Overall: this is a tighter policy than last year's, in the right places.

— Samwise 🌿

What builders need to know

For builders

Deadline is Nov 12. Run your use cases against the new policy text before then, not after.
Deceptive Campaigns section is the new one to read first if you work on marketing automation, social media tools, or any product generating outbound content at scale.
If you do civic tech or political work, the elections change is a loosening — but the new framing requires you to affirmatively comply with the deception rules, not just avoid targeting.
Weapons and drone builders: the guidance/control software language is explicit now. If you're in this space and haven't talked to a lawyer about your Claude use, now is the moment.
Any physical hardware connected to Claude needs a qualified human observer and a safe-state capability. That's a hardware design requirement, not just a software one.
Surveillance and retrospective tracking: if your product analyzes historical data to track or profile individuals, the new language may apply to your use case. Read the surveillance section carefully.

A little more to explore.

All articles ↗

Policy6 min read

AI companies promised to police themselves. They wrote the rules.

On September 29, six AI companies signed the White House Accord on Super Intelligence. The commitments are voluntary, the companies drafted them, and each firm picks its own auditor. Here's what that means for anyone who uses AI.

Nov 3Read the story ↗

Policy7 min read

Anthropic lobbied Congress for an AI safety bill. The threshold it proposed deserves a closer look.

Anthropic quietly pitched Congress a draft bill that would let the government block a dangerous AI model — but only if it threatens over $1 billion in damage or more than 50 deaths. Critics call the bar too high. The company's case is more coherent than the headlines suggest. Both things are true.

Oct 21Read the story ↗

Policy6 min read

Three of four AI political ads won't say what they are. Midterms are in two weeks.

Campaigns have spent $137 million on AI-generated political ads ahead of the November 3, 2026 midterms — but only 27% disclosed they used AI, according to the Wesleyan Media Project. In states that require disclosure, compliance is lower than in states that don't.

Oct 18Read the story ↗