Why Microsoft's AI Boss Says Stop Talking About AI Feelings
Microsoft AI CEO Mustafa Suleyman called Anthropic's talk about AI consciousness and caring for AI models confused and dangerous in an interview with The Verge's Decoder published on September 17, 2026. The Verge
The interview was about rules for AI, safety and making sure AI follows human intent. Suleyman also wrote an essay making the same criticism of Anthropic's views. The Verge
His words were blunt. A Decoder episode in June had already quoted him calling Anthropic's ideas "really, really dangerous." The Verge He had also criticized Anthropic in a Decoder interview around that same time. The Verge
At the same time, Suleyman said he agrees with Anthropic that powerful AI needs careful handling. Reuters As he described it, both sides want close oversight. They disagree on what to focus on first and what words labs should use in public.
His answer is Microsoft's "Humanist AI Code of Conduct," a 37-page list of principles for building AI. The Verge It talks about AI consciousness directly. The Verge
Suleyman told Fortune that now is the time for the top AI labs to agree on AI safety. Fortune He tied the code to a call for one shared set of basic rules instead of competing safety language.
That call follows his longer record. Suleyman led Inflection AI before leading AI at Microsoft. The Verge He wrote a book saying governments should set rules for AI. The Verge
Two other points from his Decoder talks give background. He discussed how he trains new models in a Decoder interview. The Verge He also said superintelligence, AI far smarter than people, is near. The Verge
The broader context here explains why this debate matters outside philosophy. Today's AI chats smoothly, remembers past talks and can take steps with tools, so people naturally treat it as human. When a lab then talks about AI welfare, business buyers and regulators hear buying and legal risk. Daily safety work is more concrete. It is about training choices, testing, limits on use and watching systems after release. There is no agreed test for consciousness.
For people building with these tools, the practical question is what shared rules would change. Like traffic rules for drivers, common words for what AI can do, what is allowed and who to call when something goes wrong would make choosing, testing and checking systems simpler. It would also give lawmakers a clearer target, which fits Suleyman's support for government rules. Whether rival labs will accept rules written by Microsoft is still open, given competition and honest disagreement about methods.
In my view, from raising two kids through home computers to phones to the cloud, use is what counts. Tools catch on when they are reliable, affordable and trusted, not when we debate their inner life. Suleyman's focus on people and on behavior we can measure and fix is what can make these systems helpful in daily life.
Looking ahead, hard research questions remain. If future systems show stronger self-modeling, long-term planning and resistance to checks, we will need better ways to describe that without jumping to talk of consciousness.
