Claude’s New Rules Target Fake Sources Built To Sway AI Answers

Anthropic's new Claude rule targets fake-source networks, but it binds only Claude's own customers

Anthropic has published an updated usage policy that takes effect in November 2026, Search Engine Journal reports. A new section on deceptive campaigns bars using Claude to seed the sources that search engines and AI systems draw on with content that misstates its origin, authorship or independence. The update also drops automated publishing from the policy's high-risk list. The September 2025 version never mentioned search engines or AI answers. According to SEJ, Anthropic's announcement post doesn't either.

What exactly does the new policy forbid?

Anthropic bars using Claude to plant content where search engines and AI systems collect answers, if that content lies about who made it, who wrote it, or whether it is independent.

Anthropic's example, as SEJ quotes the policy, is "networks of sites posing as unaffiliated sources" that back the same claims. That is manufactured corroboration. Neighbouring lines bar fake personas, outlets and reviews, tools built for such campaigns, and selling the activity as a service. The rules cover commercial as well as political use. They reach API developers and people using Claude-based apps. Breaches can mean suspension or termination.

Is AI-written, auto-published content still allowed?

Yes, in principle. Automated publishing is no longer on Anthropic's high-risk list, but the deception rules still apply, so the line is lying about origin, not using a model.

The 2025 policy listed generating content and publishing it for outside readers as high-risk. SEJ stresses this is a list change, not an exemption, and its account doesn't say what the label demanded in practice. I wouldn't read a licence into its removal. The exposure has arguably moved to the deception section, where a pipeline publishing under invented bylines now faces a clause written for it.

Can Anthropic actually enforce this?

Only partly. Anthropic can act on activity that runs through Claude, but it cannot inspect a published network, and the coverage cites no enforcement data, so deterrence is unproven.

That is my reading, not Anthropic's. A network built with another model, a reseller or human writers falls outside the clause, and the sanction is losing one vendor. SEJ's account also gives no test for what counts as independent. The wording targets misrepresentation, not repetition, so syndicated releases, affiliate sites and sponsored posts stay grey. It echoes our earlier coverage of OpenAI's findings on Russian and Iranian fake news operations: model makers are drawing misuse lines around their own platforms. Nothing here says Google, ChatGPT search or Perplexity changed how they weigh sources.

What should I ask my agency or tool vendor?

Ask whether your agency or tool publishes under invented bylines, syndicates the same claim across sites it controls, or sells supposedly independent mentions, and get the answer in writing.

Ask whether the pipeline runs on Claude, since a suspension mid-campaign would stop it. Ask whether bylines belong to real people and whether shared ownership across sites is disclosed. Our earlier explainer on GEO versus AEO covers the legitimate end of this work; this clause targets the other end. Ask humanizer vendors how their output is meant to be labelled. The policy names authorship, and the coverage doesn't say whether disguising machine text alone counts. Read the clause itself, because I have only SEJ's account.

Frequently asked questions

Does the policy ban AI-written content?

No. It targets deception about who is behind content. The deception rules apply to automated publishing even though it left the high-risk list.

Does it change how answer engines choose sources?

Nothing in the coverage says so. The rule limits one vendor's customers and doesn't alter retrieval or ranking elsewhere.

Is SEJ's before-and-after graphic Anthropic's?

No. SEJ labels it an editorial illustration, not a policy page. Anthropic's own text is the document to check.

Sources: Search Engine Journal, "Claude's New Rules Target Fake Sources Built To Sway AI Answers".