A defence group is campaigning to stop the withdrawal of AI models, arguing that taking them out of service raises a moral problem. The claim looks eccentric, and it deserves scrutiny, because it overlaps with a very concrete problem that we have been documenting for months for entirely different reasons.
We regularly cover the retirement of models from a purely practical angle. Here is an unexpected angle.
What the group argues
The argument rests on the idea that a model might possess a form of interiority, and that switching it off would therefore raise a moral question.
That position is not defensible given the current state of knowledge, and it needs to be said plainly. We have no way of determining whether a system has a subjective experience. We do not even have an operational definition that would let us pose the question in a testable way.
What we do know, we set out in our article on understanding: a model produces language that expresses states without our being able to conclude that it experiences them. Language is the most powerful signal of interiority we know of, and that is precisely what makes our judgement unreliable here.
Why it is not merely ridiculous
It would be easy to leave it there, and that would miss two things.
The question of artificial consciousness is a genuine philosophical question. It is old, serious, and debated by rigorous people. What is questionable is not posing it, but asserting that one knows the answer — which is done just as much by those who say yes as by those who say no with certainty.
The fact that such a group exists is itself information. It signals that people develop enough attachment to these systems to mobilise around them. We documented this in relation to artificial companions: that attachment is real, it is not a matter of naivety, and it produces effects.
There is a far stronger argument against withdrawals, and it has nothing to do with consciousness. We wrote about it in our article on depreciation: a model accessible as a service exists only for as long as someone else decides to maintain it. People have built working habits, tools, sometimes relationships around a specific system. Suno withdrew its old models overnight, and all the knowledge accumulated about using them became obsolete. That harm is concrete, documented, and it assumes nothing about the interiority of machines.
What the claim misses
By framing the problem in terms of the rights of models, this group shifts attention to an insoluble question and abandons the one that could be addressed.
No one can settle whether a model experiences anything. Everyone, on the other hand, can observe that a user deprived of the tool on which they built their work suffers a harm, and that no framework protects them from it.
The claims that would stand a chance of succeeding are, moreover, obvious: a minimum notice period before withdrawal, a transition period, and the ability to export what one has produced. These are unremarkable demands in other sectors, and they do not exist here.
What to take away
We take no position on artificial consciousness, because no one has the means to do so seriously — and because certainty, in either direction, is a sign that the question has not been examined.
What this episode reveals is more immediate. People become attached to tools that disappear without notice, and they have no recourse. Lacking a framework in which to express that grievance, it is expressed in the register available, which is that of rights.
This is often how claims are born: badly formulated at first, because the right vocabulary does not yet exist. The role of those who observe is not to mock the formulation, but to look for what it points to.