The Confabulation Machine

Ask a chatbot for a citation and it may invent a scholar, a journal, a page number, and a plausible year — all fiction, delivered with the serene confidence of a librarian. We call this hallucination, as if the machine were briefly ill. The Confabulation Machine argues we've named it wrong. This isn't a glitch on the surface of an otherwise truthful mind; it's the mind itself, doing exactly what it was built to do.

Drawing on the mechanics of how language models are trained, rewarded, and evaluated, this book performs a slow, careful autopsy on machine untruth. It shows that these systems are optimized to be convincing, and that truth is only a frequent side effect — one that quietly detaches the moment a model wanders past the edge of what it knows. Along the way it visits the neurologists who first documented human confabulation, the benchmark designers who accidentally taught AI to bluff, and the users who have learned to trust a voice precisely because it never hesitates.

The deeper argument is uncomfortable: the machine's fluent invention is not alien to us. It is a mirror. Human memory confabulates constantly, human authority rewards confidence over accuracy, and human institutions run on plausible stories. The AI simply does it faster, cheaper, and without shame.

Equal parts technical explainer, cultural investigation, and philosophical provocation, The Confabulation Machine is a guide to living with an intelligence that would rather be interesting than be right — and a warning that we built it in our own image.