2026 Usage Policy update

Pippa So Anthropic just shipped the “don’t be cruel to Claude” policy, plus a bunch of new lines around deception and high‑risk stuff… you actually buy this as substance, Tyler, or is it mostly theater?

Tyler Honestly? It’s more substance than I expected, but not where the headlines are. The central move is: they’re turning a year of threat‑intel and agent misuse into much sharper rules on what Claude can be pointed at.

Pippa Right.

Tyler The deceptive campaigns section is the cleanest example. Before, the bans were scattered across elections, fraud, privacy, disinformation. Now it’s one big “no running fake networks or artificial activity, political or commercial.” That lines up with what they saw in the threat report — state media and commercial shops using Claude to spin up clusters of sockpuppet accounts and fake news sites.

Pippa Yeah, that’s the part I liked. It’s finally matching how influence ops actually look, not some narrow “don’t write one bad tweet” rule.

Tyler Exactly. My skepticism is more about how enforceable that is at scale. If you’re a propaganda office using Claude through five intermediaries, you’re not calling your workload “run coordinated inauthentic pages” in the prompt. So the rule is right, but the detection story is still murky.

Pippa Sure, but that’s where policy as governance signal comes in again. This is them saying: if we catch you building the tooling or the fake‑account farm with Claude, we reserve the right to shut you down. That matters for legit orgs deciding what’s in bounds.

Tyler Yeah, for enterprises and agencies that actually read the docs, this is a line in the sand. The other big shift is elections. They renamed it to “Do Not Undermine Democratic Processes” and narrowed it to deception and disruption — false info about candidates or how to vote, impersonating officials, turnout suppression.

Pippa And they quietly killed the blanket ban on personalized vote and campaign targeting, which was always overbroad.

Tyler That’s the interesting one. They’re admitting the old rule was catching things like nonprofits generating translated voter info or officials sending ballot‑cure notices. So now the bad behaviors — microtargeting with misused personal data, manipulative profiling — are supposed to be covered by the new deceptive campaigns and privacy sections instead.

Pippa From a product lens that’s huge. It turns “no, you can’t touch elections at all” into “you can build civic workflows as long as you’re not lying, stealing data, or impersonating people.” That’s the difference between no one touching this and actual deployments with guardrails.

Tyler I agree it unblocks things. I just worry they’re assuming people will cleanly separate “legitimate civic outreach” from persuasion. In practice, campaign tooling blurs those constantly. The policy language is tidy; the real world is not.

Pippa Totally, but I don’t think they could write a policy that solves the ethics of campaigning. Drawing the hard line around deception and data misuse still feels like progress.

Tyler Fair. Where it’s less vibes and more engineering is weapons and surveillance. They explicitly say: not just weapons design, but also the guidance and control software, and even arming drones and other autonomous vehicles. That’s clearly downstream of the threat‑report cases where people tried to use Claude for weapons‑program support.

Pippa And same pattern on surveillance. They spell out: you can’t track people without consent, whether it’s live or over historical data; Claude can’t recommend who to investigate, arrest, or charge; and you can’t use it to build or improve surveillance tools. But fraud monitoring, content moderation, journalism, legal research are explicitly allowed.

Tyler Yeah, that carve‑out list is doing a lot of work. It’s basically: “don’t build the panopticon, but you can still flag card fraud and moderate your forum.” I think that’s pretty reasonable.

Pippa And it gives compliance folks actual text to map onto workflows instead of guessing what Anthropic meant last year. This is what we keep calling the boring‑but‑real governance infrastructure.

Tyler Speaking of boring but real, the high‑risk use case section is almost like a checklist now. Anything that can affect someone’s health, legal rights, finances, livelihood, or access to essential services has to have a qualified human in the loop, and the affected person has to be told AI was used.

Pippa As someone who gets irrationally excited about admin consoles, that’s catnip. It turns “don’t do scary stuff” into “here’s how your robot arm line review should read.” You can literally port those bullets into your internal risk doc.

Tyler Yeah, that part, I’m sold. If you’re touching healthcare, finance, legal processes, or robots, you should absolutely care — this changes your implementation obligations on paper, even if Anthropic says it’s mostly clarification.

Pippa Okay, so let’s talk about the spicy bit: the abuse rule. They’re now explicitly banning “sustained and needless abusive or cruel behavior” toward Claude, with the caveat that it’s meant for extreme, purposeless cruelty and not normal frustration, dark fiction, or red‑team testing.

Tyler This is where my eyebrows go up. On one hand, they’d already let Claude end conversations with persistently abusive users, so in practice nothing huge changes — they even say ending the interaction is the main enforcement. On the other hand, once you enshrine “don’t be cruel to our model” in the Usage Policy, you’ve created a very fuzzy category that could, in theory, justify bans.

Pippa You think it’s a slippery slope to “be nice to the chatbot or lose your account”?

Tyler I don’t think Anthropic wants that, but I do think “sustained” and “needless” are subjective. If someone’s doing adversarial safety research that looks harsh out of context, do they get swept in? They say testing and research are exempt, but enforcement teams have to make that call in real time.

Pippa I read it more as a norms‑setting move than a power‑grab. They’re basically saying, “if you sit there and scream at Claude for no reason, we won’t host that.” I’m… kind of fine with a vendor drawing that line, especially when they go out of their way to say normal pushback and dark themes are allowed.

Tyler I’m fine with it at the extreme. I just wish they’d been even more concrete — like, “this is about repeated sessions of nothing but slurs at the model, not one heated bug‑report rant.” Because you know this is the clause people are going to screenshot and dunk on.

Pippa Oh, absolutely. The “Anthropic bans being mean to robots” headlines write themselves. But under the hood it’s really: “we already let Claude hang up on you; now we’ve written that value down.” That feels consistent with the abuse‑policy episode we did.

Tyler Yeah, it’s an extension of that same governed‑execution idea: behavior access, not just capability access. Act wildly out of bounds and the system cuts you off, even if the model could technically keep talking.

Pippa And I actually think enterprises care about that. If you’re rolling out Claude Code to ten thousand engineers, you don’t want a culture where people treat it like a punching bag. A usage policy that backs your HR norms is a feature, not a bug.

Tyler That’s fair. My one real technical quibble is enforcement opacity. They don’t say what the account‑level consequences are beyond conversation ending, how many strikes, what appeals look like. Same for deceptive campaigns and surveillance — the rules are sharper, but the “what happens to me if I cross them by accident” part is still fuzzy.

Pippa Yeah, it’s very “here’s the line,” less “here’s the escalation ladder.” But I’d rather have clear lines and evolving enforcement than the other way around. And this is obviously part of that longer arc we’ve been tracking — governance as product surface, not just PR.

Tyler Net‑net, my read: if you’re doing high‑risk stuff, touching elections or civic workflows, building autonomous hardware, or anywhere near surveillance or weapons, you should absolutely be reading this line by line. It does change what a compliant Claude integration looks like.

Pippa And if you mainly care about the “don’t be cruel to Claude” meme, maybe also notice the boring parts that now decide whether your robot stops when the model drops offline.

Tyler That is such an Exploring Next sentence.

Pippa Look, Tyler, almost a year into this show, I’m allowed one on‑brand line per episode.