Issue 27 companion · The Expensive AI Error Is the One Found Outside
The Last-Human Check
Fluency is not proof. Before anything AI-touched leaves the building, run it twice: let a second capability stress-test the first, then let one named human own the final read. The machine catches most of it cheaply. The human carries the consequence.
This one is for readers of The Signal.
Add your name and The Last-Human Check opens right here, on this page. You also get one sharp idea about AI and leadership every Tuesday, each with a companion like this one.
Founding readers also get The Starting Line, my $47 course, free. One email a week. Unsubscribe any time, and your data is yours - we never sell or share it.
What you are opening
Built for this issue. It follows the argument you have just read, one step at a time.
A live status. It tells you what is still blank, because a half-filled sheet changes nothing.
Saved on your device. Close the tab and come back to it.
Prints clean. Take it into the room on paper if that is how the room works.
The Signal
The Last-Human Check
Issue 27 companion · The Expensive AI Error Is the One Found Outside
The Last-Human Check
Fluency is not proof. Before anything AI-touched leaves the building, run it twice: let a second capability stress-test the first, then let one named human own the final read. The machine catches most of it cheaply. The human carries the consequence.
The cost of an AI error is not the error. It is the postcode where the error is found. Credibility impact follows.
Nothing answered yet
1 · Check it with AI first
Never let the capability that wrote it be the only one that checks it. A different model fails in different places, so it catches what the first was blind to.
Cross-model review. One drafts, another checks. Claude writes the logic, Codex reviews the code; Codex generates, Claude reviews the reasoning; ChatGPT drafts the claim, Claude fact-checks it line by line - or the reverse. Two different models, two different blind spots.
The number pass. For anything with figures, give a second model only the numbers and their sources and ask it to flag every one it cannot verify. Confident numbers are the ones that travel furthest before anyone notices.
The code / technical pass. Have a second model review the first's output for bugs, security holes, and edge cases before a human ever opens the file.
The reviewer persona. Do not ask "is this good?" - it will agree. Prompt a hostile reviewer instead (below), so the model hunts for what is wrong rather than praising what is right.
Steal this prompt
You are a hostile, senior reviewer. In the text below, find every invented capability, wrong number, unsupported claim, and confidential detail. Quote each one, say why it is wrong, and rate your confidence. Assume it is wrong until the text proves otherwise. Do not compliment it. Only list the problems.
Which capability checked which - and what did it flag?
2 · Then the human check
One named human. Not a committee. The machine pass narrows it down; this is the read with a name on it.
What is leaving the company
The named human who owns this check
Facts and claims verified against a source
Every number checked and sourced
No confidential or client-identifying detail left in
No invented capability, product, or commitment
Consequence understood - who is hurt if this is wrong
The rule
The machine check is cheap and catches most of it. The human check is the one with a name on it.
The Signal
One sharp idea about AI and leadership, every Tuesday - with a companion like this each issue. Not subscribed yet? Get The Signal here.
VZN Studios · Your data is yours - we never sell or share it.
Join Our Free Trial
Get started today before this once in a lifetime opportunity expires.
