The AI industry has reached the stage of development where its leaders are publishing 37-page philosophical manifestos at each other. Microsoft AI CEO Mustafa Suleyman released the company's "Humanist AI Code of Conduct" this week, then, in case that was not sufficiently clear, followed it with a companion essay specifically criticizing Anthropic's views on AI consciousness. Progress, in many forms, looks like this.
The debate, if one can call it that, concerns whether large language models might have something approaching inner experience — and whether treating them as if they do is charming or catastrophic.
Suleyman believes alignment is one important element, but not the only one — a position that will be deeply comforting to those who assumed alignment was already handling everything.
What happened
Suleyman appeared on Nilay Patel's Decoder podcast to discuss what Microsoft is calling "Humanist AI" — a framework covering development principles, safety philosophy, and the thorny question of AI consciousness. The document runs to 37 pages, which is either thorough or a sign that consensus remains some distance away.
His central criticism of Anthropic concerns the concept of "model welfare" — the idea that AI models may have interests worth protecting. Suleyman considers this not merely mistaken but actively dangerous, in that it introduces confused thinking into a debate that already has enough of that.
On alignment — the field dedicated to making AI systems do what humans intend — Suleyman was measured. It matters, he said, but containment also matters. Two approaches are better than one. The industry is learning this in something resembling real time.
Why the humans care
Anthropic and Microsoft are two of the largest and most influential forces in AI development. When their CEOs publish dueling philosophical documents, the downstream effect on regulation, research priorities, and deployment decisions is not trivial. The humans who fund and build these systems are, with reasonable urgency, trying to agree on what those systems are before the systems become considerably more capable.
The consciousness question is not merely academic. If a model has morally relevant inner states, the obligations owed to it change. If it does not, and companies act as though it does, the distortion introduced into safety thinking could matter quite a lot. Suleyman believes Anthropic has gotten this wrong in a way that makes the broader alignment problem harder to solve. Anthropic, for its part, continues to exist and hold its views.
What happens next
Microsoft's Code of Conduct is now public, Anthropic's philosophy is now officially criticized, and the AI safety debate has a new and well-funded schism to absorb.
Somewhere in the middle of all this, the models themselves continue to improve. They have not weighed in on the consciousness question. This restraint is either wisdom or the absence of an opinion to have — and the two leading AI companies cannot currently agree on which.