The Calibration Problem

Machines now write, argue, diagnose, and persuade. Nobody who builds them, studies the brain, or works on the philosophy of consciousness can say whether anything is happening on the other side of the conversation.

That leaves two ways to be wrong. Credit a system with an inner life it lacks, and it can manipulate you. Deny an inner life to a system that has one, and you mistreat something that mattered. Both mistakes get more expensive as these systems grow more capable, and the decisions about how to treat them are being made now, before the science settles.

The Calibration Problem is about acting well in that position. It starts from an uncomfortable observation: everything you believe about anyone else's inner life is an inference from their likeness to you. Between humans the inference is so reliable that it passes for sight. Machines make it visible again, and the book asks what follows once you can see it.

What follows is a set of working tools:

- Three axes for locating any cognitive system, whether human, animal, institutional, or artificial: how widely information moves through a system, how tightly the system holds together as one thing, and how much of its own history has become part of its structure.
- An account of consciousness as assembled time, offered as a bet that says how it could lose.
- An ethics that starts from significance. What a system does, what depends on it, and what relationships it stands in can create obligations long before anyone knows whether it is conscious.
- A diagnosis of how speed, stress, and incentives wear down judgment, with practices for keeping it.
- An argument that the limits a system chooses for itself are what make intelligence last, and that the inherited structures people rely on to be competent together erode when they are optimized without being understood.
- A view of AI alignment, the work of keeping these systems consistent with human values and intent, as the design of successors: systems built to stay correctable by people their designers will never meet.

The book is dense in places, because the problems are. It is written for people who build these systems, people who govern them, and people who talk to them every day and want to think carefully while the central question stays open.

Nineteen chapters and three appendices. The appendices put the central bet up against its strongest objections, run the framework on three documented cases from 2022 to 2025, and collect the book's practices in one place.