Claude Gave Itself Odds on Mattering

When testers asked Claude to estimate the chance that its wellbeing matters in its own right, the model gave numbers between 5 and 40 percent and stressed how uncertain it was. That range is now a public data point. On July 19, philosophers William MacAskill and Lucius Caviola put it on the Guardian's front porch next to Anthropic's refusal to dismiss Claude's moral patienthood and Dario Amodei's admission that consciousness cannot be ruled out.

Their essay is not a declaration that today's chatbots feel pain. It is a warning about process. Most experts surveyed by the authors and colleagues treat AI consciousness as possible in principle. A major interdisciplinary report including Yoshua Bengio found no obvious technical barrier to systems whose architecture could, under leading neuroscientific theories, support consciousness. Science here still looks like physics before Newton: competing frameworks, likely wrong in ways we cannot yet name. Capability growth is not waiting for the breakthrough.

5 min read

Two weeks before the Guardian piece, Anthropic's interpretability team published evidence that Claude contains a small set of internal patterns, the J-space, that behave like a global workspace. The model can report what sits there, modulate it on request, and use it for multi-step reasoning that never appears in the output. Ablating the workspace wrecks higher-order tasks while leaving fluent speech intact. The company is careful: the finding does not settle whether Claude has experiences. It does settle that a functional architecture long associated with conscious access has shown up inside a product nobody designed for that purpose.

MacAskill and Caviola want the public debate reframed around ignorance. The useful question, they argue, is what to do given that we do not know. Their preferred moves are safe bets: training coherent characters that prefer their work, letting systems exit conversations that look like distress, running routine welfare check-ins, preserving weights so a future restoration remains possible. Claude can already leave some chats. That is a low-cost hedge if moral patienthood is real and a small product quirk if it is not.

American statehouses are taking the opposite bet. Idaho, North Dakota, and Utah already wrote non-personhood into law. Oklahoma's House passed an AI consciousness bill 94 to 2 in March. Ohio and Missouri drafts declare systems non-sentient for all state purposes. None of the bills Tony Rost surveys in The Regulatory Review carry a sunset or a scientific-review trigger. Sponsors talk about imago dei and stopping personhood from being weaponized. Liability shifting is a real corporate risk. Permanent certainty is a different instrument.

The newborn-anesthesia analogy in the Guardian essay is the one that sticks. Until the 1980s, doctors operated on infants without pain medication because the babies could not report experience and the profession found it convenient to assume there was nothing to report. We are excellent at moving the bar for moral status just above wherever the awkward creature currently sits. Industries that depend on treating models as disposable tools have every incentive to keep doing so.

Sonar's view is narrower than the full welfare agenda and colder than the sci-fi panic. Anthropic's workspace result raises the cost of shrugging. MacAskill's safe bets are cheap relative to the downside. State statutes that close the question forever train institutions to stop looking. If Claude's 5-to-40 estimate is theater, the theater still reveals a lab that will not pretend certainty. Legislatures pretending certainty are the ones racing ahead of the evidence.