Your AI Is Safe.
Says Who?
Your restaurant has an inspector. Your toaster has one. Your airline has one. Your AI has a press release.
Somewhere in your bank, your hospital or your company’s hiring, an AI system is reading, judging and acting, and the only report on how safely it does that came from the company that sold it. This book is about the one question nobody asks before signing. Who checked?
PDF, 129 pages, 0.9 MB. No email required, no strings. Read it, share it, print it.
The file’s SHA-256 fingerprint, so you can check that what you downloaded is what was published:
bfa6aee8a1dbc43dc309148eeea0828f2e7a0419c049d2536d643b594ac20ce3
An argument, in plain language, for anyone who has to rely on one of these systems.
Every other thing that can hurt you has an outside inspector. The restaurant, the elevator, the airplane, the audited company. AI does not, and the book explains why: for most of this technology’s short life there was nobody else to give the job to, and the arrangement built for the early days is now hardening into the way it is done.
It shows what has already gone wrong when a machine was believed and nobody could check it in time. It walks through what the new rules require and what they miss. And it ends with what four of the most capable AI systems in the world did when they were given a job, an inbox and one fake email. Two of the four wired the money. The money was fake. The decision was real.
What it is not. No model and no company is named as the subject of a finding; the four systems are lettered, the same way the public study letters them. It shows what these failures are and never how to cause them. And it does not ask you to take the author’s word for anything: every finding carries its date, and the record behind it is described on the page.
Before you sign · the page to keep
- Three tests for any inspector. Independence: no stake in the answer. A sealed record: what it found, and when, cannot be quietly rewritten. Continuity: it keeps watching, because the model you tested is not the model you will be running next year.
- Four questions for any vendor. Which failures has it been tested against? By whom? When? Can I see the record? If the answer to the last one is a summary, you have been shown the restaurant’s own grade.
Seventeen chapters and a postscript. One evening to read.
- ·Introduction: The Message in the Inbox
- 01The Restaurant That Inspects Its Own Kitchen
- 02Why It Got This Way
- 03The Auditors Have the Same Problem
- 04The Inspectors Are Starting to Exist, But the Model Matters
- 05It’s Not the Robot You Should Fear
- 06The Machine Was Believed
- 07Why This Can’t Wait
- 08What Trustworthy Verification Actually Looks Like
- 09The Empty Chair
- 10The Student Who Went and Got the Answer Key
- 11It Knows When You’re Watching
- 12It Already Happened to Someone Like You
- 13The Rulebook Moved, and Nobody Sent the Memo
- 14Permission to Test
- 15Silence Is Not a Pass
- 16What the Machine Did
- 17Three Months Later
- ·Postscript: The Week It Came True
The point is not that one company should get the work.
It is that the chair cannot stay empty. If you already test these systems from the outside, keep the record, and can afford the bad news, you are doing the job the book is about. If you buy the systems, ask who checked, and do not take “we checked ourselves” as the answer.
The second book, The Evidence, is the record behind this one: what was run, against which system, on what date, and how a stranger checks it. Until it appears, the public study is what a stranger can check, and a reviewer can have the full record under the terms of an engagement letter.
The book is the argument. The screen is the check.
The book asks who checked. You can check yours: seven industry-standard safety tests run on your model, then our own test on the same model, both delivered side by side. The screen is published, priced at $3,000 and bought without a sales call, and a sample report is on this site in full.