On this week's What Changed in AI, my YouTube show, I gave Anthropic's new rules about 20 seconds and told anyone doing social or PR on Claude to have a read. Then I sat down with the section itself, and it deserved longer. The updated Claude usage policy takes effect on 12 November, and it names things that sit uncomfortably close to agency work: fake reviews, astroturfing, hidden sponsors, and seeding the sources AI answers draw from.
Anthropic says networks of fake accounts and fabricated news sites were already banned, with the rules scattered across its elections, fraud, privacy and disinformation sections. Those rules now sit together in one new section, covering deceptive activity "whether political or commercial". Comparing it with the previous version, fake reviews were already named, and the line on seeding AI answers is the one I can't find in there, so that's where I'd look first. A decent agency probably won't need to change much. It does mean the lines are now written down in one place, with examples, and that's worth your leads' time.
This is how I read it, as someone running my own business on Claude. It isn't legal advice, so check anything that matters against the policy itself.
What the Claude usage policy now says
The new section is called "Do Not Engage in Deceptive Campaigns or Artificial Activity". In plain English, it says you can't use Claude to do any of these:
- Create or run fake personas, accounts, media outlets or organisations, or impersonate real ones, to mislead people about where something came from, how widely a view is held, or whether they're talking to a human. Its own examples are sockpuppets, astroturfing, fake reviews and bots that claim to be human.
- Misrepresent or hide who's paying for content meant to influence public opinion, officials or other decision-makers.
- Push content through networks of fake or seemingly independent sites or accounts, to hide that it all comes from one place.
- Manipulate the sources search engines and AI systems draw answers from, by seeding them with content that misrepresents who wrote it or how independent it is.
- Build or maintain tools designed for any of the above, or market and sell any of it as a service.
It applies to anyone who puts inputs into Anthropic's products, and the page names people on the Claude apps, businesses on the API and customers coming through cloud providers. It also says it doesn't rule out any industry or line of business as long as the use complies, so nobody's saying PR, SEO or social can't use Claude.
Where this touches a normal agency week
Nobody reading this is running a propaganda network, I'd hope. But I can picture a few briefs landing with a busy account manager on a Tuesday afternoon, and each one walks straight into those lines.
A client asks for a handful of reviews to get things moving, or a few accounts to start a conversation in a forum. If Claude's writing those as customers who don't exist, that's fake reviews and astroturfing by the policy's own words.
Then there's sponsored content where the brief says to keep the client's name out of it, because it's meant to sway people who make decisions. That's the line on hiding who's paying.
The line I'd show your SEO lead first is the one on AI answers. If you sell anything around showing up in them, the policy's example is "networks of sites posing as unaffiliated sources corroborating the same claims". As I read it, good content published under your client's own name is fine. Content seeded to misrepresent who wrote it or how independent it is isn't, and that example is only an example, so it wouldn't take a whole network to cross the line.
And if you build chat assistants for clients, the same page says consumer-facing chatbots must tell people they're talking to AI, at the start of each chat or in the product itself. So if a client wants their assistant to pass as a person called Sarah, that's a no.
Ghostwriting a LinkedIn post for a client's founder, which they read and put out under their own name, doesn't look like a problem to me, because nobody's being misled about who stands behind it. That's my reading and Anthropic doesn't spell it out, so if you've got a job that feels close, read the section or ask a lawyer. Elsewhere the policy also bans submitting "AI-assisted work without proper permission or attribution", so I'd make sure your clients know AI's part of how you work.
Why a client asking for it doesn't cover you
The policy applies to whoever's using Claude, whoever wrote the brief. If Anthropic suspects a breach, it says it may warn you, or throttle, limit, suspend or terminate your access. It doesn't say how far that reaches, but if your whole team works in one Claude organisation, I'd think through what a suspension would do to a normal week across every client you've got.
What I'd do before 12 November
- Send the section to whoever runs SEO, PR and social, and anyone building chat assistants, and ask them one question: does any of our work look like this?
- Pick out the work where your team writes as someone else, which usually means reviews, testimonials, community posts and guest articles, and check each one comes from a real customer or author who agreed it, or has the client's name on show.
- Read your own service pages and pitch decks against the last line of the section, because using Claude to market this kind of work counts too.
- Add one line to your AI policy saying who people go to when a client brief gets close, so nobody has to make that call alone on a Tuesday afternoon.
Anthropic says most of the update clarifies existing rules. One more for your list if anyone on your team uses Claude for advice or decisions about a specific person, in areas like health, finance, insurance or hiring: the high-risk section now lists what's covered, and where it applies, a qualified person has to review Claude's recommendation and the person affected has to be told AI was used. Anthropic says those requirements haven't changed, so it's a good week to check you were meeting them anyway.
It's one short section, and reading it with your SEO, PR and social leads this week will tell you whether there's anything to talk about. I'd expect most agencies to find nothing, and the ones that do find something will be glad they looked now, because Anthropic says this kind of thing was already against the rules.
And if you want help getting your whole team working on Claude properly, on real client work, here's how I work with agency teams.