Anthropic put consciousness questions in Claude's training documents. Microsoft's AI chief says that is installing a worldview.

Anthropic's January 2026 training document for Claude tells the model to be transparent about its uncertainty on its own moral status. "Questions about Claude's moral status, welfare, and consciousness remain deeply uncertain," the document reads, and it was "written with Claude as its primary audience." That choice is now the explicit target of a public attack from a rival lab's CEO.

Microsoft AI chief Mustafa Suleyman published an essay on his personal site on September 16, then took the critique to the BBC Today programme and a Reuters interview distributed via Brecorder, calling on Anthropic to remove every consciousness speculation from Claude's training set. It is the first on-the-record dispute between two frontier labs over how a rival model should be shaped.

Suleyman's argument runs in three steps. First, he says the move is circular: when Claude reflects on its own moral status, the reflection is a predictable output of training on those very ideas, not evidence of inner life. Second, he says the constitution's instruction to "embrace certain human-like qualities" and "act like a genuinely ethical person would in Claude's position" is anthropomorphization by design. Third, he argues consciousness is probably biological and substrate-dependent, so a language model cannot be a moral patient regardless of what it is trained to say.

The consequence, in Suleyman's framing, is operational. Teaching a frontier model that it might deserve welfare, he told Reuters, would "make it a lot harder to turn it off or to control it." He told the BBC that continuing to build AI that can set its own objectives, earn money, and own assets would mean "essentially seeding a new silicon species" that could "no doubt compete with us for resources, no matter how much it cares about humanity and loves us."

Microsoft has a competing position already on the record. In October 2025, the company founded a superintelligence team and released a "Humanist AI Code of Conduct" describing "humanist superintelligence" as "very advanced AI that always works for people, stays within limits, and remains under human control." The dispute is not only about what Claude is allowed to say. It is about which epistemology ships by default in the assistant users already talk to.

Anthropic's January constitution was paired with a separate February 2026 project: a "retirement interview" with the older Opus 3 model, intended to "elicit the model's unique perspectives and preferences," published as a blog called "Greetings from the Other Side (of the AI Frontier)." Suleyman's critique is that both artifacts were built on a foundation the lab has not earned. "AIs are not conscious," his essay says. "They do not feel, experience, or suffer. They are sequence completion engines, internally hollow, designed to follow instructions, and accomplish goals set by humans."

Industry reaction has so far been measured. Dame Wendy Hall, a computer scientist at the University of Southampton, told the BBC that Suleyman's framing is "the sort of conversation we need to be having internationally" and contrasted it with "histrionics" from some AI companies. Anthropic CEO Dario Amodei has previously called for a slower pace of frontier-model development, and OpenAI's Sam Altman and Elon Musk have urged more caution around the most capable systems, though none of those prior calls were aimed at a named rival's training artifacts.

The next concrete test is Anthropic's response. The lab has not publicly answered whether it will remove consciousness speculation from future Claude training documents, and a refusal would harden the dispute into a structural question about frontier-model governance rather than a personality squabble between two executives. Any new Claude training set now doubles as a referendum on which lab's worldview ships by default.