Microsoft AI Chief Challenges Anthropic on Consciousness and Safety
Microsoft AI CEO Mustafa Suleyman called Anthropic's philosophy on AI consciousness and model welfare 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, the work of keeping AI behavior in line with human intent. Suleyman also published an essay making the same criticism of Anthropic's views on consciousness and model welfare. The Verge
The language was direct. 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 in a Decoder interview in that same period. The Verge
At the same time, Suleyman said he shared Anthropic's focus on safely managing AI. Reuters As he described it, both sides agree that frontier systems, the most capable models, need careful oversight. The dispute is over what that oversight should prioritize and what language labs should use in public.
His alternative centers on Microsoft's "Humanist AI Code of Conduct," a 37-page statement of principles for AI development. The Verge The document addresses AI consciousness directly. The Verge
Suleyman told Fortune that now is the moment for leading AI labs to unite around AI safety. Fortune He presented the code as part of a call for a common baseline across labs, rather than competing safety vocabularies.
That call follows 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 give technical background. He discussed his approach to training new models in a Decoder interview. The Verge He also said superintelligence, AI that would far exceed human ability, is near. The Verge
The broader context here helps explain why this dispute carries weight beyond philosophy. Labs now ship systems that invite people to treat them as human by design, through fluent conversation, persistent memory and agentic behavior, meaning AI that can take multi-step actions with tools. When a lab then talks about welfare or moral status, business buyers, developers and regulators hear buying and legal implications. Day-to-day safety work is more concrete. It lives in training choices, evaluation design, deployment controls and monitoring after release. Consciousness and welfare have no agreed tests, thresholds or response plans, and the concern is that focus on possible sentience could pull review time away from concrete failures like data leakage, tool misuse, over-permissioned agents and weak evaluations.
For teams building on these platforms, the practical question is what a shared code would change. Common terms for capabilities, limits, acceptable use and escalation paths would simplify model choice, red teaming, or deliberate adversarial testing, and audit work. It would also give policy teams a clearer target to regulate, which fits Suleyman's support for government involvement. Whether rival labs will adopt principles drafted by Microsoft is still open, given competition and honest disagreement about alignment methods.
In my view, shaped by watching my two children grow up through the shifts from desktop to mobile to cloud, the focus on use over inner life makes sense. Technologies take hold through reliability, cost and trust, not through claims about consciousness. Suleyman's humanist framing keeps attention on human agency and measurable system behavior, things engineers can test and operators can enforce. That is what could make powerful systems genuinely useful.
Looking ahead, that pragmatism still leaves a research gap. If future systems show more consistent self-modeling, long-horizon planning and resistance to oversight, the industry will need better instruments to describe those traits without defaulting to consciousness language.
