Suleyman Calls Anthropic's Consciousness Stance Confused and Dangerous, Presses Shared Safety Rules
Microsoft AI CEO Mustafa Suleyman criticized Anthropic's philosophy around AI consciousness and model welfare as confused and dangerous in an interview with The Verge's Decoder published on September 17, 2026. The Verge
The interview focused on AI regulation, safety and alignment. Suleyman also published an essay criticizing Anthropic's philosophy around AI consciousness and model welfare. The Verge
The language was blunt. A Decoder episode in June had already carried his description of Anthropic's speculation as "really, really dangerous." The Verge He had also criticized Anthropic during an interview with Decoder in that same period. The Verge
At the same time, Suleyman said he shared Anthropic's focus on safely managing AI. Reuters That distinction matters. The disagreement is not about whether frontier systems require careful oversight. It is about what oversight should prioritize, and what language labs should use in public.
The immediate vehicle for Suleyman's alternative is Microsoft's "Humanist AI Code of Conduct," a 37-page statement laying out principles around AI development. The Verge The document addresses AI consciousness directly. The Verge
Suleyman told Fortune that now is the moment for top AI labs to unite around AI safety. Fortune The comment connects the code to a broader call for a common baseline across labs, rather than competing safety vocabularies.
That call fits his longer record on regulation. Suleyman was CEO of Inflection AI before becoming Microsoft's AI chief. The Verge He wrote a book arguing that governments should regulate AI. The Verge
Two other threads from his Decoder conversations help explain the technical side of his position. Suleyman discussed his approach to training new models in an interview with Decoder. The Verge He also said superintelligence is near. The Verge
For practitioners, that combination is familiar. Training choices, evaluation design, deployment controls and post-deployment monitoring are where safety work lives. Consciousness and welfare, by contrast, lack agreed measurement methods, operational thresholds or incident response playbooks. Suleyman's critique appears aimed at keeping attention and engineering effort on the former.
The broader context here is worth spelling out, and in this author's view it explains why this dispute has traction beyond philosophy. Labs now ship systems that invite anthropomorphism by design, through fluent dialogue, persistent memory and agentic behavior. When a lab then discusses welfare or moral status, enterprise buyers, developers and regulators hear something with procurement and liability implications, even if researchers intended an exploratory ethics discussion. Worth flagging is the risk that loose talk about sentience could distort safety priorities, shifting review time toward speculative harms and away from concrete failure modes like data leakage, tool misuse, over-permissioned agents and brittle evaluations.
Looking at what this means for teams building on these platforms, a shared code has practical appeal. Common terminology for capabilities, limitations, acceptable use and escalation paths would simplify model selection, red teaming and audit work. It would also give policy teams a clearer surface to regulate, which is consistent with Suleyman's argument for government involvement. The open question is whether rival labs will sign onto principles drafted by Microsoft, given competitive pressure and genuine methodological disagreement about alignment.
My own bias, after watching my children grow up alongside each platform shift from desktop to mobile to cloud, is to discount metaphysical debates and focus on use. Technologies settle into daily life through reliability, cost and trust, not through claims about inner life. In that sense, Suleyman's humanist framing is pragmatic. It centers human agency and measurable system behavior, which are things engineers can test and operators can enforce.
That pragmatism does not resolve the underlying research problem. If future systems exhibit more consistent self-modeling, long-horizon planning and resistance to oversight, the industry will need better instruments to describe those properties without defaulting to consciousness language. Until then, Suleyman is drawing a line that many practitioners will recognize. Build for human benefit, measure what you ship, and keep speculative ontology out of production safety discourse.
