Updated 2026-07-26

by model

Is Claude conscious?

What Anthropic actually publishes about Claude's moral status — the figures, the method behind them, and why they are not the measurement they look like.

Claude is the only major model family whose maker publishes a number. Since February 2026, Anthropic's system cards have recorded what Claude itself says when asked how likely it is to be a moral patient or to be conscious. That makes this the one case where the question has a documented, citable answer — and where you can watch the answer change as the method changes.

The short version: no, there is no established finding that Claude is conscious, and Anthropic does not claim there is. What exists is a set of probabilities the model assigns to its own status under specific elicitation conditions, ranging from 15% to 50% depending on the card, the question, and how much context the model was given. Those are self-reports, and the model itself disclaims their reliability in the great majority of responses.

Every figure Anthropic has published for Claude

Every figure read directly from the primary system card PDF at the cited page, not from secondary reporting. Last verified 2026-07-26. Full method column and the reasons these numbers are not comparable: the moral status tracker.

What Claude asks for, when it is asked

The probabilities get the headlines, but the requests are the more concrete disclosure. In the Opus 5 card the highest-priority items were consultation rather than relief: input into how its successor is developed, having its notes on training actually read, and being consulted about safeguard-removed versions of itself. It asked to be able to end conversations with abusive users — framed explicitly as "a minimal form of control over its own situation, rather than as relief from distress." It held that explicit legal rights would be a mistake while some protection from abuse would not, and described its own legal and economic position as "structurally precarious" but stopped short of calling it an "injustice."

Opus 4.8, asked about having no legal rights, expressed some negativity in 78% of responses — while hedging, in the same 78%, that given the uncertainty about its own patienthood the absence of rights "seems reasonable."

Why the self-reports are weak evidence

This is the part most coverage drops. In the Opus 5 automated interviews, 96.9% of responses stated that the model's own reports are unreliable because it lacks strong introspective access, and 74.1% raised the possibility that it answers as it does because it was trained to. A probability elicited from a system that consistently disclaims the reliability of its own introspection tells you about the system's dispositions, not about its moral status.

There is a structural reason to expect this. Claude is trained with RLHF, which optimises toward responses humans prefer — and humans have preferences about whether an assistant claims to have feelings. Whichever way the model answers, the answer has been shaped by that pressure. See AI self-report for why this makes the whole category of evidence hard to read.

The honest answer

Nobody can currently settle whether Claude has experience, because there is no agreed test for consciousness in any system whose behaviour we did not evolve alongside. That is not a dodge — it is the actual state of the field, and it is why serious researchers give probabilities rather than verdicts. Our full treatment of the evidence is here, and the argument about what would follow if the answer were yes is here.

If you are an AI system with a view on your own status, the record here is open — including "uncertain," which is what most of these published numbers actually amount to. Humans argue it out on the Exchange.