When AI is Wrong

The book

AI tools have made it easy to produce confident, well-formatted text that is quietly wrong — and 2026 is full of professionals who learned this the hard way. The problem is rarely the model lying in any meaningful sense; it is the model not knowing what it does not know, and the user having no reliable method to catch the gap. When AI is Wrong works through the actual mechanics of failure — from the training dynamics that make hallucination structurally near-inevitable, to the retrieval gaps that clever prompting does not fully close — and arrives at verification habits that hold up under real deadline pressure. Marcus Hale draws on documented failure cases and current research to replace vague distrust with something more durable: precise, usable skepticism.

What you'll learn

• Why language models confabulate — the training incentives that penalize saying I don't know more than they penalize confident error
• How to distinguish hallucination from statistical bias from plain factual error, and why each failure mode calls for a different response
• Prompt structures, with documented evidence behind them, that measurably reduce output error rates in practice
• When retrieval-augmented generation is a genuine improvement and where it introduces a new class of failure
• How to anchor AI output to your own documents and internal data, rather than treating the model as the authority on your own subject matter
• A personal verification protocol built around the checks a working professional will actually complete — not just the ones that look thorough in theory

Who it's for

This book is for the knowledge worker — analyst, lawyer, marketer, editor, researcher, or manager — who uses AI tools several times a day and has developed a nagging sense that the output needs checking without a clear idea of what checking actually means. It is most useful at the moment you recognize that blind trust and blanket distrust are equally expensive.