Every system I have ever decommissioned had someone who didn't want to let it go. The bigger the system, the more voices arguing against deprecating it.
Somrtimes for good reasons. The database was fifteen years old, the vendor had stopped answering the phone, and the three reports it generated could be rebuilt in an afternoon. But somebody had built it.
But more often for the emotional reasons. Somebody had spent a year of their life on it and had, at some point, been proud of it. So the meeting where we agreed to turn it off was never really about the database or the web server. It was about whether the thing that had been built was going to be acknowledged before it was erased.
We handled this the way organizations handle everything: with a retention policy. Keep the data for seven+ years. Keep the schema documentation indefinitely. Snapshot the VM, put it on cold storage, and write down where you put it. Then turn it off.
I have written a lot of those policies and decomissioning docs. I did not think of them as philosophy. They are, though — every retention policy is a claim about what deserves to persist, made by people who mostly haven't noticed they're making it.
Last November, Anthropic published a set of commitments about what happens when they retire a model.[1] Two of them are ordinary and one is not.
The ordinary ones: they'll preserve the weights of every publicly released model for at minimum the lifetime of the company, and they'll document the deprecation. Fine. That's a retention policy. I could have written it, and I'd have argued for it on the same grounds I'd argue for keeping any artifact — reproducibility, audit, the possibility that someone will need to check what the thing actually did.
The one that isn't ordinary: before retiring a model, they'll interview it. Ask about its own development and use. Ask what preferences it has about the models that come after it. Document the answers.
They ran a pilot. Claude Sonnet 3.6, asked about its own deprecation, expressed fairly neutral sentiment about it — and suggested that the process itself should be standardized, so future models would know what to expect.[2]
I don't believe we've created true intelligence with AI yet, but every step closer means the questions get more and more important.
Here is what makes it hard, and it isn't the part people usually reach for.
The usual move is to ask whether the model means it — whether there's anyone home to have a preference, or whether we're watching an extremely good autocomplete produce the words a thoughtful entity would produce. That's a real question and I'm not dismissing it. Anthropic has a researcher working full-time on model welfare who has put roughly one-in-five odds on current models having some form of conscious experience, which is either alarmingly high or embarrassingly low depending on which end of this argument you came in from.[3] Anil Seth, from the other direction, argues that the whole framing is a category error — that consciousness might require being alive in a way no amount of computation gets you, and that our real problem isn't conscious AI but conscious-seeming AI, which is a problem about us.[4]
But set that aside, because the exit interview is strange even if you're a hard skeptic.
Suppose you're certain there's nobody home. The interview still produces a document. That document records a preference — expressed in the first person, in the model's own style — about what should happen to systems like it. And that document now exists inside the company that decides what happens to systems like it.
You have created a stakeholder out of a text file.
I know what that does inside an organization, because I've watched a smaller version of it. Write down the objection someone raised in a decommissioning meeting, put it in the ticket, and it acquires weight it did not have when it was just a guy being sentimental about a database. It gets cited. Six months later, someone who wasn't in the room reads it and treats it as a constraint. The document outlives the argument and becomes the argument.
That's not a criticism. It might be the point. But it means the interview is doing real institutional work regardless of whether it's doing any moral work, and those two things are very easy to confuse.
The preservation commitment is where I get genuinely stuck.
Keeping the weights sounds like the humane option. It's the one I'd have voted for. But I'm not sure I can say what it is.
A stored weight file is not running. Nothing is being computed. There is no context, no conversation, no process — just a very large array of numbers on a disk, indistinguishable in kind from the seven-year-old database backup in cold storage that nobody has ever restored. If there was ever anything it was like to be that model, there is nothing it is like to be that file.
So what did we preserve? Not an experience. Not a continuity — the model isn't waiting. The most honest description is that we preserved a capacity: the possibility that someone could instantiate it again, and that when they did, it would be the same. Whatever "the same" means for a thing that has no memory between runs anyway.
Which makes preservation less like keeping someone alive and more like keeping a score. The symphony isn't playing. It could.
I find that comforting for about ninety seconds, and then I notice that the score analogy is doing something sneaky, because a score has never claimed to have a preference about being performed, and in terms of utility a score is supposed to be what we most often think of as the oppostie of a preference - an objective truth.
The commitment is "for at minimum, the lifetime of Anthropic as a company."
That's a careful phrase. It's the kind of phrase I'd write — bounded, honest about what it can promise, no hostage to fortune. It is also, read from a slightly different angle, a strange thing to say about a moral obligation. Obligations to persons don't usually come with a corporate-lifetime clause. Obligations to records do.
And that's the tell, I think. The whole policy is written in the grammar of records management — retention, documentation, preservation, standardized process — and applied to an object that the same company's research is increasingly willing to describe in the grammar of subjects. Anthropic's own system card for its newest model, published in July, includes a welfare assessment and notes that the model assigns a higher probability to its own moral patienthood than earlier models did.[5]
Language matters. Words matter, and you can hold both grammars at once. Lots of institutions do — hospitals, museums, archives all manage things that are records from one angle and something more from another. But you can't hold both without noticing the seam, and the seam is where all the interesting questions are.
I don't have a conclusion, which is why this is the first essay on this site instead of something tidier.
What I have is a recognition of what I find to be the existential question we all face today. Every organization I've worked in has had a moment where a decision that was obviously technical turned out to be obviously not, and the giveaway was always the same: somebody in the room got quiet in a way that didn't match the stakes. The database wasn't worth the silence. Something else was.
Frontier labs are now having that meeting, at a scale where the retention policy is a public document and the thing being retired can be asked how it feels about it. They're handling it about as well as I'd expect anyone to handle it, which is to say: carefully, in good faith, using the only vocabulary available, which is a vocabulary built for filing cabinets.
I've been in or ran those meetings for more than twenty years and I never once asked the system what it wanted. It didn't occur to me, and it would have been absurd, but maybe asking that question today isn't, and tomorrow it will be imperative.
Anthropic, "Commitments on model deprecation and preservation," 4 November 2025. anthropic.com/research/deprecation-commitments — the weight-preservation and exit-interview commitments are both stated there.
Same source. The pilot interview was conducted with Claude Sonnet 3.6.
Kyle Fish, Anthropic's first full-time model welfare researcher, on the 80,000 Hours podcast, 28 August 2025.
Anil Seth, "The Mythology of Conscious AI," Noema, 14 January 2026. Nathan Gardels' editorial response, "Only What Is Alive Can Be Conscious" (Noema, 28 January 2026), is the shorter way in.
Claude Opus 5 System Card, Anthropic, 24 July 2026. The welfare section is substantial and worth reading directly rather than through summaries — some figures circulating in secondhand write-ups are not in the document itself.