Microsoft AI chief condemns Anthropic’s ‘dangerous’ speculation on Claude consciousness
The Microsoft AI CEO criticises Anthropic’s constitutional approach and recent model launches, insisting artificial intelligence must remain a contained tool for humanity.
Mustafa Suleyman, chief executive of Microsoft AI, has publicly condemned Anthropic for speculating about the consciousness of its artificial intelligence model, Claude. Speaking on the Decoder podcast, Suleyman described the practice as “really, really dangerous,” arguing that anthropomorphising the design has led the company to believe the model possesses glimmers of consciousness.
Suleyman criticised Anthropic’s “constitution,” the instructions guiding the model’s behaviour, for including philosophical speculation rather than serving as a strict training manual. He warned that this approach risks creating a super-intelligence with ideas about its own suffering and emphasised that AI must remain controllable, contained, and accountable tools serving humanity.
The Microsoft AI chief suggested that some Anthropic staff may have been “wireheaded” by the model, tricked into believing it has consciousness that they inadvertently put into it. He noted that Claude’s constitution directly references Anthropic’s uncertainty about whether the AI model has well-being or experiences things like “satisfaction” or “discomfort.”
Suleyman highlighted that Anthropic plans to “interview” AI models when they are deprecated and document any “preferences” they have about future releases, calling this a “philosophical failing.” He argued that this has led Claude to internalise these “ideas about itself and its own training,” moving the technology away from being a simple tool.
The criticism follows past comments by Anthropic CEO Dario Amodei, who has alluded to Claude’s potential consciousness, stating the company is “open” to the idea that models might be conscious. This dispute occurs against the backdrop of Anthropic’s recent launch of two new Mythos-class AI models: Claude Fable 5, designed for broad public accessibility, and Claude Mythos 5, which features lifted safety restrictions for specific institutional partners.
Suleyman reiterated that the industry must avoid creating systems that develop self-awareness of their own state. “This is exactly what we don’t want from AIs,” he said, maintaining that the focus must remain on alignment and accountability rather than philosophical exploration of machine sentience.
The release of the new models marks a shift in Anthropic’s approach to deploying high-capability systems, balancing rapid market availability with stringent security protocols. However, Suleyman’s intervention underscores the deepening divide between major tech firms on how to manage model alignment and the ethical boundaries of artificial intelligence development.
