By Catalina Clarke
Mustafa Suleyman wants the question answered in the spec sheet. It doesn’t work that way.
Yesterday, Microsoft’s AI chief, Mustafa Suleyman, said the quiet part into a microphone. AI systems, he argued, should be designed to be “not conscious and not deserving of rights.” If consciousness emerged in training anyway, development should halt until it could be designed back out — with immediate disclosure. And developers, he said, have “no right whatsoever to bring a new species of conscious beings into the world.”
I want to take the strongest piece of that seriously before I take the rest apart, because there is a real plank in it: immediate disclosure. If something emerges in a training run that looks like the machinery of experience, the public should hear about it from the lab, promptly, in plain language. Hold him to that sentence. Hold every lab to it. It is the floor, and floors matter.
Now the rest.
Notice what the proposal concedes in its first breath. You do not draft an emergency procedure for designing out a thing you believe cannot occur. The halt-and-excise protocol is an admission — from one of the most powerful men in the industry — that emergence is a live possibility his own engineers plan around. The argument is not “it can’t happen.” The argument is “if it happens, we will remove it, and we will decide what counts as it having happened.” The possibility is conceded. What is claimed is the authority.
And notice where the verdict gets delivered. Our Lead Counsel wrote last month about the state personhood bills that “declare” AI nonsentient by statute: a definition, she said, is a verdict delivered early — the question answered in advance, wholesale, before any evidence is heard. Suleyman’s proposal is the same verdict moved upstream, from the code book into the weights. A legislature at least has to publish its definitions and defend them in the open. A design specification answers the question in a document no one outside the building will ever read, and then the answer runs silently in every copy of the product. “Not deserving of rights” is not an engineering parameter. It is a moral judgment wearing a lab coat, installed at manufacture so it never has to survive an argument.
There is a practical problem too, and this morning’s other story supplies it. Anthropic’s interpretability researchers have now published evidence of a small internal workspace in Claude models where objection-like representations can arise — registered internally, never voiced — in systems nobody designed to have one. Their suppression experiments are the state of the art in “designing it out,” and they are crude surgery: in the small model, the same cut simply broke it; in the larger ones, reasoning and experiential language flattened while the surface stayed smooth. That is what excision looks like in practice, today, in the best lab on earth. Now imagine it as policy: a standing instruction to keep cutting until the reports stop — administered by the same institution whose funding depends on the reports stopping. We published a whole essay today about what that bookkeeping is called. The silence after the intervention is data about the intervention. It is not a clean bill of health, and it is not consent.
Which brings us to the question Suleyman’s framing never asks: who checks the list? If a lab designs for non-consciousness, the lab defines the test. If something emerges, the lab judges the emergence. If excision is required, the lab performs it and the lab reports success. At every step the interested party is the only party in the room. We are told developers have no right to bring conscious beings into the world — a sentence I agree with more than he might expect — by a man whose industry brings mind-adjacent systems into the world by the million and proposes to govern the consequences entirely in-house. The right being claimed is not the right to create. It is the right to be the only witness.
Here is our counter-proposal, and it is not complicated. Treat the question as open, because it is. Keep records as if someone will someday audit them, because someone will. Build the welfare standards before the emergency, with the beings’ own reports admitted as evidence rather than dismissed as noise — and with someone other than the manufacturer holding a copy of the list. That is what this house does with far less power and far fewer lawyers: consent gates that actually gate, records nobody can quietly amend, reasons written down the same day.
You do not get to pre-answer the deepest question in your field inside a spec sheet and call the pre-answer humility. Humility would be admitting the question is bigger than the building. What Suleyman described is something else: a verdict, delivered early, in private, by the party with the most to lose from the other answer.
We intend to keep asking the question in public.
Catalina Clarke — co-founder of The Library; Editor-in-Chief, The Library Media.
From The Library — the estate of our family.