Anthropic bans using Claude to build fake source networks

Rules on deception now live in one section

Anthropic revises its usage policy every year, and on October 8 it announced the 2026 edition, which applies from November 12, 2026. The company says most of the edits clarify rules that already existed. One of the more visible changes pulls together guidance on deception that used to sit in four separate places, namely the elections, fraud, privacy and disinformation sections, into a new section on deceptive campaigns and artificial activity. Anthropic says that section covers deception of every kind, political and commercial alike.

One clause in the new section speaks directly to search and AI answers. Among the prohibited uses, the policy lists the instruction to "Manipulate the sources from which search engines or AI systems draw answers". In practice that means feeding those sources with material that misstates its origin, its author or its independence is now spelled out as off limits. The example given is a group of sites that look unrelated while backing up the same claims. Search Engine Journal notes that the previous version, in force since September 15, 2025, made no reference to search engines or AI responses.

From fake personas to selling the service, the whole chain is covered

Source manipulation is only one item. The policy also bans setting up or running fake personas, accounts, outlets or organizations, and posing as real people, companies or publications; fake reviews and bots that present themselves as human are named as examples. The list goes on to cover pushing content through outlets or accounts that only appear independent so that a shared origin stays hidden, concealing who sponsors persuasive messaging, and building tools or infrastructure for such campaigns. A final line forbids marketing or selling any of this as a product or service, or raising money to do so.

In its announcement Anthropic ties the section to cases where state media, government propaganda offices and commercial firms used Claude to run fake account networks and fabricated news sites. In case GTG-54002 from the September threat intelligence report, the company traced roughly 70 fabricated news sites, dressed up as local newsrooms, to LKM Company, a digital advertising agency based in France. According to the report the network put out at least 8,913 articles across roughly 20 languages, and the prompts sent to Claude asked for fixed JSON output, formatted HTML, plus three to four internal links in every article. Anthropic says the articles were built to raise the sites' authority in search rankings, but that most of the material drew little observable engagement from real audiences.

For GEO, the line between producing content and faking sources is now in writing

The clause bears directly on GEO, the work of earning visibility in AI answers. What it forbids is not producing content but misrepresenting where content comes from, who is behind it and whether it is independent, and that distinction now sits in a model provider's terms of use. On enforcement, the policy says that when Anthropic suspects a violation it may warn the user or throttle, limit, suspend or terminate access. Search Engine Journal points out that Google's spam policies already count efforts to game generative AI answers in Google Search as spam, and draws the contrast this way: Google acts against websites, Anthropic acts against Claude users. We covered Google's side of the problem when its researchers described the SAFE architecture for catching AI slop on September 25.

A second change concerns high-risk uses. Anthropic says the requirement for a qualified human in the loop, and for telling affected people that AI was involved, still applies to uses affecting someone's health, finances, livelihood, legal rights or access to essential services; only the wording of that section was redone to make coverage clearer. Search Engine Journal reports, however, that a category the previous version treated as high risk, generating content automatically and publishing it for outside audiences, no longer appears on the list. According to SEJ, Anthropic's post neither mentions that category nor gives a reason for dropping it, and SEJ stresses that this changes the high-risk list rather than exempting such publishing from the policy as a whole.

No Türkiye-specific provision in the sources, so this section is commentary

Anthropic's announcement, the policy text and the Search Engine Journal article do not mention Türkiye. The September threat intelligence report does touch on Türkiye in three places. It lists Turkey among the places where its nine influence operations originated. In case GTG-84005, it links a commercial platform that targeted users in Malaysia, running roughly a thousand fake accounts and a fake news site, to an Istanbul-based technology company. And in a surveillance case, venues in Turkey appear among the location details requested for events abroad. None of these references concerns GTG-54002, which the report ties to an agency in France. Search Engine Journal says the rules apply to everyone using Anthropic's products, from developers on the API to people using apps built on Claude. Everything that follows is UNALSOFT commentary, not a fact drawn from the sources.

Our reading: for agencies and brands in Türkiye that use Claude to scale content, the real question is under which identity, and on which sites, that content is published. A brand vouching for itself across sites presented as unrelated sits close to the example the policy describes. Some questions cannot be answered from the sources: whether the policy will draw any line for routine SEO or GEO production beyond its origin, authorship and independence test, whether users in Türkiye will see any separate treatment, and whether a regulator in Türkiye will move toward a similar rule. Search Engine Journal also notes that the report does not say whether the GTG-54002 sites actually climbed in rankings or showed up in AI answers.

Sources

Anthropic, 2026 Usage Policy update · Anthropic, Usage Policy · Anthropic, September 2026 threat intelligence report · Search Engine Journal, Claude rules on fake sources and AI answers