The new Anthropic usage policy takes effect on 12 November 2026, and one of its new sections names growth tactics outright: sockpuppets, astroturfing, fake reviews, and bots that claim to be human. If Claude sits anywhere in your marketing stack, you have nine days to check what you have built.
Anthropic published the update on 8 October and says most of it clarifies rules that already existed. That is fair for the weapons and surveillance sections. It is not fair for the part a marketer runs into: a new standard with six specific prohibitions.
Rules on deceptive activity used to be scattered across the elections, fraud, privacy and disinformation sections. They are now one standard with its own heading, bringing the total to fifteen Universal Usage Standards in the policy text.
The policy covers anyone who submits inputs, spelled out as individuals using the apps such as Claude.ai and Claude Code, developers and businesses on the API, customers reaching Claude through cloud providers and authorized resellers, and the end users of products that integrate Claude. If a client’s chatbot runs on your Claude key, both of you are inside it.
Anthropic says the rewrite targets misuse it has already tracked, including state media, propaganda offices and commercial firms running networks of fake accounts and fabricated news sites, as MacRumors reported on the day of publication.
Six prohibitions sit in it. Creating or operating fake personas, accounts, media outlets or organizations. Concealing who sponsors content meant to shape public opinion. Distributing content through networks of fake or ostensibly independent outlets that hide a shared origin. Manipulating the sources that search engines and AI systems draw answers from. Building or maintaining tools and infrastructure that enable any of it. Marketing, selling or raising money for those services.
The rule attaches to the tactic, not the motive. Agency work rarely frames itself as deception: a fake reviews problem arrives as a seeding plan, a sockpuppet arrives as a “community account”. The examples are blunt enough to settle those arguments internally before 12 November.
One bullet is new ground for marketers: manipulating the sources from which search engines or AI systems draw answers. The mechanism Anthropic describes is seeding those systems with content that misrepresents its origin, authorship or independence, with the example of sites posing as unaffiliated sources that happen to back the same claims. That is a direct hit on one strain of answer-engine optimisation.
The line is independence, not publishing. A well-sourced article on your own domain, attributed to your own brand, remains ordinary content marketing. Five review domains with invented bylines pointing at one client is now a named prohibited use, and so is building the tooling that stands them up. Using Claude to write a post, draft a hundred ad variants or run a brand voice pass is untouched.
Consumer-facing chatbots and AI agents that communicate directly with external users must disclose that users are interacting with AI rather than a human. The mechanics are specified: the disclosure has to appear at least at the start of each chat session or in the product interface, and you do not have to name Anthropic, Claude or the model to satisfy it. For most readers that is one sentence in a DM auto-responder and one line in a website widget greeting.
One sentence in it costs more to act on. Users are responsible for ensuring that agents they build or deploy comply with the policy, including what those agents do through tools, browsers and connected systems. An agent that posts, replies or signs up for things on your behalf is your compliance surface, not Anthropic’s.
Separately, under the fraud standard, the policy prohibits generating content for spam and submitting AI-assisted work without permission or attribution.
The two high-risk requirements have not changed: a qualified person must meaningfully review a high-risk AI recommendation and have authority to change it before it is delivered, and the affected individual must be clearly told that AI was used to produce it. What changed is that the Anthropic usage policy now enumerates eleven areas instead of gesturing at categories: legal, medical, finance, credit, insurance, housing, employment, education and credentials, healthcare access, public benefits, and legal status.
Under finance that includes a personalized recommendation to buy, sell, hold or allocate among specific investment products. Under legal it includes advising an individual on their rights, obligations or liability. If you write client copy in those verticals, the question is whether a deliverable is a recommendation aimed at an individual.
The exclusions are the useful half. General education and wellness content sit outside the requirements. So does internal drafting and analysis, explaining a decision a qualified person already made, applying fixed rules, and business operations not directed at specific individuals. A post about how mortgage rates work is education. A letter telling one named applicant what to do is not.
It does not require you to label ordinary AI-assisted creative work. The disclosure duties are narrow: chatbots that talk to external users, and synthetic media presented as coming from a real candidate, official or outlet.
It does not govern your other tools. ChatGPT, Gemini, your scheduler and your ad platforms each have their own terms. It also does not replace the law: the policy states that its requirements do not replace local legal requirements and restrictions, so a stricter rule in your market still wins.
Two changes run the other way. The elections section, now called Do Not Undermine Democratic Processes, dropped its blanket ban on personalized vote and campaign targeting, although targeting built on deception or misused personal data is still barred elsewhere. The new rule against sustained, needless cruelty toward the models is enforced by Claude ending that conversation, not by anything happening to your account, and it does not cover ordinary frustration, pushback, dark creative themes or testing.
The one that can quietly catch an agency is the supported regions policy, also effective 12 November. It excludes people while physically located in an unsupported region, entities incorporated or headquartered in one, and entities majority-owned or controlled, directly or indirectly, by persons or entities there. If you use offshore contractors, ask that question this week rather than next quarter.
No. The disclosure requirements cover consumer-facing chatbots and politically framed synthetic media, not ordinary AI-assisted writing. Platform rules and advertising codes may still ask for labels, and those are separate obligations.
The terms you operate under change, so a tactic that was buried in a general clause becomes a named prohibited use. Anthropic has not published a dated enforcement schedule alongside the update, and the only mechanism it describes in detail is Claude ending abusive conversations.
The blanket prohibition was removed from the elections section. That is not the same as permission: deceptive targeting and the misuse of voters’ personal data remain prohibited under the deception and privacy standards.
Three of the six checklist items take under an hour each. The fourth, answer-engine tactics, is the only place where a working practice may have to stop. Read the policy against what you actually run, not against what you would describe in a pitch.
This is the kind of thing we work on with clients at Social Lady. If you want a second opinion on your AI automations before 12 November, get in touch.
