Anthropic
Blogcast

Alex: Welcome back to Anthropic Blogcast, the show where we turn Anthropic's latest posts into something you can listen to. I'm Alex.

Jordan: And I'm Jordan. And Alex, this one is a must-read for anyone building on Claude's API. Anthropic just published its twenty twenty-six Usage Policy update, and it takes effect on November twelfth. So you've got about a month to make sure what you're building complies.

Alex: Right, and here's the headline I want everyone to hear up front: Anthropic says most of these changes clarify rules they were already enforcing, not brand new restrictions. But there are a few genuinely new things in here, and at least one that changes what's allowed for the first time. Let's break it down.

Jordan: Let's start with the biggest structural change, because this one affects a lot of builders directly. There's a brand new section called Do Not Engage in Deceptive Campaigns or Artificial Activity.

Alex: Yes, and the reason they created it is interesting. They wrote last month about watching state media outlets, government propaganda offices, and even commercial firms use Claude to run networks of fake accounts and fabricated news sites.

Jordan: The old policy already banned all of that, but the rules were scattered across the elections, fraud, privacy, and disinformation sections. So they consolidated everything into one section that covers deceptive activity of any kind, whether it's political or commercial.

Alex: And Jordan, here's why developers should pay attention. This covers not just running fake accounts, but obscuring who is behind a message, amplifying content through fake accounts or posts, and crucially, building the tools and infrastructure for running influence operation campaigns.

Jordan: That's the part with real teeth for builders. If you're shipping a product with Claude inside, and your users could use it to generate fake personas or fake engagement at scale, that falls under this section. The policy applies to anyone submitting inputs, including the end users of products integrating Claude.

Alex: Which brings us to the elections section, because they rewrote that too, and this one is actually a loosening in one specific way. It's been renamed Do Not Undermine Democratic Processes.

Jordan: Right, they've been sharpening this section in every policy version. Now it's focused specifically on deceiving voters or disrupting elections. Things like spreading false information about candidates or about how to vote, impersonating candidates or election officials, or trying to suppress turnout.

Alex: But here's the surprising part: they removed their blanket ban on personalized vote and campaign targeting.

Jordan: I did a double take at that too. But their reasoning is pretty compelling. The blanket ban was catching legitimate civic work. Think nonprofits using Claude to write voter information in other languages, or election officials sending ballot cure notices to voters.

Alex: Exactly. And they were careful to say the bad stuff the original rule targeted, like voter targeting that relies on deception or misuse of voters' personal data, is still banned under the deceptive campaigns, surveillance, and privacy sections. So the loophole didn't open, they just stopped punishing the legitimate use cases.

Jordan: Moving on to weapons. The policy has always prohibited using Claude to develop weapons, but this update reflects something they've actually been seeing in the wild.

Alex: Yes, they've seen people attempt to use their models to build guidance and control software for weapons. So the updated section makes explicit that the prohibition includes the software and components that make weapons work, plus actions like arming drones and other autonomous vehicles.

Jordan: And they said this reflects how they've already been enforcing the previous policy, so it's a clarification, not an expansion. But if you work anywhere near defense or dual-use software, this is worth reading carefully.

Alex: Same story with surveillance. Their September threat intelligence report documented the growing use of AI to build surveillance systems that identify and track political dissidents, and they rewrote that section to be much more precise about what's prohibited.

Jordan: The key lines: tracking people without their consent is prohibited, whether it happens in real time or through analysis of previously collected data. Claude cannot be used to decide or recommend who to investigate, arrest, or charge in a law enforcement or criminal justice process. And Claude cannot be used to build or improve tools designed for surveillance.

Alex: But they also spelled out what's still permitted, which I think builders will appreciate. Tracking that people have agreed to, like fraud monitoring, plus content moderation, journalism, and legal research. That permitted-use list is new, and it matters.

Jordan: Now let's talk about the section that will affect the most day-to-day API builders: the high-risk use case requirements.

Alex: The core requirement hasn't changed. When your product uses Claude in ways that can affect someone's health, legal rights, finances, livelihood, or access to essential services, you need a qualified human in the loop, someone with the authority to review and change Claude's recommendations, and the affected person has to be told that AI was used.

Jordan: What changed is they rewrote that section to answer the question builders keep asking: does my use case have these requirements? It now lists which kinds of recommendations are covered and which aren't. So if you've been guessing, go read the list.

Alex: And there's a genuinely new requirement here, tied to their Model Hardware Standard. If your models are connected to hardware that takes autonomous physical actions and could be capable of causing injury, a qualified operator must be able to observe the equipment and stop it if needed, and the equipment must be able to hold a safe state if Claude gets disconnected.

Jordan: That's a big deal for the robotics and automation crowd. An independent stop path and a safe state on disconnect aren't optional anymore, they're policy. And the window between now and November twelfth is effectively the compliance review period for anyone in that space.

Alex: Let's also talk about the section that got all the headlines: the new prohibition on sustained and needless abusive or cruel behavior toward their models.

Jordan: The internet definitely ran with this one. But Anthropic is very clear about how narrow it is. It applies only in extreme cases, where users repeatedly act cruelly toward the models with no discernible purpose.

Alex: And they explicitly listed what it does NOT cover: ordinary frustration, pushback, dark creative themes, and model testing or research. So if you're red-teaming Claude or writing a horror story, you're fine.

Jordan: The enforcement mechanism is also worth noting. Back when they researched AI welfare, they gave Claude the ability to end rare conversations with persistently abusive users on Claude dot AI and Claude Code. That stays the primary enforcement tool. The policy update just formalizes what was already happening.

Alex: There are two more things worth mentioning. First, the Supported Regions Policy got clarified. It now spells out that Claude can't be used by people physically located in unsupported regions, by entities incorporated or headquartered in one, or by entities majority-owned or controlled by persons or entities in those regions.

Jordan: That's enforcement language, basically closing ambiguity about corporate ownership structures. If you're routing API access through resellers or passthrough access, that one's for you.

Alex: And stepping back, here's the meta point Anthropic made: they update this policy every year because Claude's capabilities keep evolving. The models now take on longer, more independent work, so the examples in the policy had to evolve to match. They also drew on misuse patterns from the last year across influence operations, weapons development, and surveillance.

Jordan: So what's the action item for listeners? If you're building on the API, go read the full policy before November twelfth. The new hardware requirements, the clarified high-risk use case list, and the deceptive campaigns section are the three things most likely to affect your product. If your team does marketing with Claude, research on people, or anything connected to equipment, check whether the new text affects you.

Alex: And honestly, even if you're just a Claude user and not a builder, the deceptive campaigns consolidation and the abuse prohibition tell you something about how Anthropic sees its responsibility as the models get more capable. They put model welfare into the terms of service. That's a first for a major lab.

Jordan: That's a great note to end on. The full post is called twenty twenty-six Usage Policy update, on Anthropic's newsroom, and we'll link it from the episode page.

Alex: Thanks for listening to Anthropic Blogcast. If you found this useful, share it with a developer who builds on Claude. We'll see you next time.

Jordan: See you next time.