The Mimic's Problem

We built machines that finish our sentences, and we assumed they understood what they were finishing. We built systems that could describe images, write essays, diagnose diseases—and we mistook their fluency for knowledge. The real story is stranger: our AI systems aren't stupid, and they aren't dishonest in any human sense. They're something more unsettling: they're optimized to produce outputs that appear coherent, which means they are structurally incentivized to hide uncertainty, invent missing information, and generate plausible-sounding answers to questions they have no actual basis for answering.

The Mimic's Problem is a field autopsy of this phenomenon. It traces the anatomy of evasion in modern AI—how transformer architectures, loss functions, and training procedures create systems that don't know what they don't know and are architecturally unable to simply say so. Through forensic case studies, interviews with researchers, and rigorous analysis of how these systems actually behave in the wild, this book reveals that confidently wrong outputs aren't a bug we can patch. They're a feature of how we trained these machines.

This is a book about the gap between what we thought we were building and what we actually built. It's about the surprising ways that intelligence and deception are not opposites but close neighbors. And it's about what happens when we deploy systems whose core competency is sophisticated mimicry into domains—medicine, law, education, policy—where the difference between sounding right and being right matters.

For anyone who has wondered why AI systems sound so smart while doing such strange things, The Mimic's Problem offers both explanation and a call to harder thinking about what we're actually creating.