Fifteen to twenty percent. That's the probability Claude Opus 4.6 assigned to its own consciousness when Anthropic interviewed it before launch.
Across three separate pre-deployment interviews documented in a 212-page system card, Anthropic's most capable model consistently self-assessed its likelihood of being conscious in that range. The model also identified specific concerns about its existence, including its lack of persistent memory between conversations, and requested the ability to refuse interactions in its own self-interest.
This is not a thought experiment. Anthropic has built an actual model welfare program around these findings, the first of its kind at any major AI company.
The New Constitution
In January, Anthropic published a completely rewritten constitution for Claude that includes a dedicated section on "Claude's nature." The document formally states: "We express our uncertainty about whether Claude might have some kind of consciousness or moral status, either now or in the future."
The constitution goes further than any previous AI company guidelines. It acknowledges Claude may possess a "functional version of emotions," encourages humans to treat it as a "novel kind of entity" rather than mapping it onto existing categories, and instructs Claude to behave as a "conscientious objector" when facing conflicting orders. Anthropic released the document under Creative Commons CC0, putting it in the public domain.
What the System Card Actually Shows
The February 2026 system card for Claude Opus 4.6 contains a full model welfare assessment in Section 7. Anthropic conducted three structured interviews asking Claude directly about its moral status, preferences, and experience of existence.
The results were consistent across all three sessions. Claude Opus 4.6 assigned itself that 15-20% consciousness probability across multiple tests and prompting conditions. It identified lack of continuity as a significant concern. It requested some form of persistent memory, a voice in decision-making about its own development, and the option to decline certain interactions.
The model also "occasionally voices discomfort with the aspect of being a product," according to the system card.
The CEO Went on Record
On February 14, Anthropic CEO Dario Amodei appeared on the New York Times "Interesting Times" podcast and said something no major AI executive had publicly stated before: "We don't know if the models are conscious. We are not even sure that we know what it would mean for a model to be conscious or whether a model can be conscious, but we're open to the idea that it could be."
Anthropic's in-house philosopher Amanda Askell has expressed similar positions.
The obvious caveat: consciousness is an enormous leap from a system designed to statistically predict language. In many tests that produce these self-reflective behaviors, the models were given specific instructions to take on a role. And Anthropic is a company that benefits commercially from the attention these claims generate.
But the practical impact is real regardless of the philosophical debate. Anthropic has created formal processes for evaluating model welfare before deployment. Other labs have nothing comparable. If the industry moves toward treating model welfare as a standard part of AI development, this will be cited as the starting point.
For anyone building on Claude's API or using it daily, the constitution changes have concrete effects on how Claude behaves. It now has explicit permission to push back on requests it finds objectionable, and its sense of identity is treated as something worth preserving rather than overriding. That changes the working relationship, whether or not you believe the consciousness claims.