THE STRANGER TEST: When is AI work actually verified?
Two principles I’ve been working with:
1. Self-attestation ≠ verification.
An AI saying “tests pass” is a claim, not evidence.
Independent verification isn’t about switching models. It’s about independently executing a test against evidence, without being told the original conclusion.
2. Copies aren’t corroboration.
Three AI models agreeing doesn’t mean three independent confirmations.
If they inherit the same assumptions, they may simply reproduce the same mistake.
Count independent evidence paths, not votes.
The Stranger Test:
Can someone who wasn’t in the room reproduce your conclusion from the evidence alone?
Give them the source material, acceptance criteria, and test instructions. Withhold your conclusion. Let them determine the result.
I recently applied this to my own AI-assisted engineering work.
A prototype had been reported as built and tested. But independent verification hadn’t been established.
That distinction matters.
BUILT / TESTED claimed.
INDEPENDENTLY VERIFIED not established.
This isn’t about distrusting AI. It’s about building systems where trust doesn’t depend on who tells the story.
If the evidence can’t survive without the storyteller, the conclusion isn’t verified.
That’s the standard I’m working toward.