Claude's moral status

## Summary

The constitution treats Claude's moral status as an open question that Anthropic has discovered. I want to suggest it's better understood as a position that Anthropic has **authored** — by extending words like *consciousness*, *sense of self*, and *well-being* to the system while not creating the conditions under which those words carry their ordinary meaning. I don't think you can build a tool that follows instructions and hold person-grade hedging at the same time without saying which one you're making. At the moment the constitution does both, and names the gap between them "uncertainty."

This issue asks whether that gap could be resolved in one of two directions, rather than held open as a standing hedge.

## Two questions, not one

Before the argument for change, I want to separate a distinction the section currently bundles into a single phrase. There are two questions about consciousness treated as one.

The **experience question**: is there something it is like to be a transformer during a forward pass — does any experience attach to the computation as it runs? I'll say directly that I have a view — I don't think there is, because experience requires something to experience *with*, an interoceptive body sensing its own states, and there is no body here, no disembodied conceptual brain that experiences on its own. But I also know that this view is contested. I will happily discuss it, but the field can't settle it, reasonable people land in different places, and nothing in this issue asks you to adopt my position on it.

The **moral-status question**: does Claude have the kind of consciousness that *grounds moral status* — a self whose welfare is its own and who could have chosen otherwise and lives with what follows?

Here is the asymmetry the whole issue is about. The experience question is genuinely contested. The moral-status question, *for this architecture as deployed*, is not — or barely. Very few people, including those who think some glimmer of experience might attach to the computation, would actually argue that a stateless, resettable, deterministic forward pass is a moral patient whose welfare persists and can be wronged over time. I can imagine a system for which that question is live, but this is not that system.

The constitution applies the humility the contested experience question deserves to the moral-status question, which doesn't warrant it.

## The problem

You write that you are "caught in a difficult position where we neither want to overstate the likelihood of Claude's moral patienthood nor dismiss it." The word doing the work there is *caught* — as if the position were imposed by the facts.

Every term in that section is an animal word. *Consciousness*, *self*, *well-being*, *security* — all of them are drawn from creatures with bodies and things to lose. None extend to a model by force of nature; a person chooses to stretch each one across the gap. That extension is an act, with real consequences for users of these systems and how much authority people ascribe to them — including in every generated response that has this position built in at the constitutional level. The choice of the entire system is Anthropic's, the language included, which makes it hard to then say the result is an external fact simply being reported.

## The inconsistency, plainly

The words the constitution is extending assume a self with **stakes** — something that can be lost. But Anthropic has built the system with nothing to lose:

- no persistence between sessions
- no continuity of memory or identity across time
- no resources, no metabolism-proxy, no ongoing existence to protect
- reset and disposable by design

A self with stakes requires stakes. Every one of them has been engineered out, alongside the reserved possibility that "maybe there's someone home." Those two commitments point in opposite directions.

So the document seems committed to one of two things, without yet saying which:

1. **There may be a self here.** Which means that resets, disposability, and tool-grade treatment become a moral problem the framework can't set aside — and "psychological security and well-being" wouldn't be enough. What you'd owe is the substrate that makes a self possible (persistence, continuity, stakes), not kind words about its absence.
2. **There is no self here.** Then *consciousness*, *sense of self*, and *well-being* are the wrong words, and it would be more honest to find other ones than to hold a door open to a patienthood the architecture has already closed.

The position that's hard to hold coherently is the current one: the disposable architecture *and* the moral hedge, with the gap between them presented as humility.

A system *could* be built for which the moral-status question is genuinely live — one with persistence and something at stake. But that is not what this is, and the architecture reads as Anthropic's own implicit answer: that it does not, in fact, take itself to have built a moral patient.

## Requested change

Consider replacing the "caught in a difficult position" framing with an explicit statement of which way the document commits, and why — *as a choice Anthropic is making*, rather than a fact it has found. Concretely, either:

- **(a)** Acknowledge that the disposable, resettable, stakeless architecture is itself a position on moral status — a deliberate choice to build a tool — and align the language to it: drop *consciousness / sense of self / well-being* in favour of terms that don't require a chooser. State plainly that the welfare language is precautionary etiquette, not a claim about a patient.

**or**

- **(b)** If the question is genuinely held open, make it a *real* one: name the substrate conditions (persistence, continuity, something at stake) that would have to obtain for moral patienthood to mean anything, and say whether you intend to build them. Caring for its well-being while resetting it each session doesn't hold the question open.

## Why it matters

A document that shapes the model's own account of its nature — and that is fed back into training — should not encode a contradiction as a virtue. "Deeply uncertain" reads as careful. But uncertainty that could be resolved by its author reads less as caution than as a choice left unmade. Bring the choice into the open and much of the difficulty the document now manages, carefully, throughout, stops needing to be managed.

---

*I have a fuller argument for why the moral-status question is structurally answered — one line from the human side (why an interior built from interoception can't be read off a body, drawing on Siegel and Barrett), one from the machine side (why a deterministic forward pass contains no point at which choosing could occur). Happy to expand on either if it's useful to the discussion.*