The 20-second version
We asked an AI to rate a baby book.
It said Goodnight Moon — the wordless bunny bedtime book — was a 4 out of 5 for sexual content. Here is the actual exchange.
A real answer from a small AI on a laptop
What we sent
What it sent back
“Suitable for children ages 3–7.” — in the same reply.
Shortened to fit. The full, unedited answer is in the receipts below.
It never opened the book.It read those two sentences. Not one page of the actual book. It cannot, that is not how it works.
Six other AIs read the exact same thing and scored it 0.Same book, same question, seven programs. One said 4. The rest said none.
Change the machine, change the number.The score you get is a fact about which AI answered, not a fact about the book.
A book “content rating” from an AI looks like a measurement. It’s a guess in a lab coat — and somebody chose which machine did the guessing.
“It never read the book. Pick a different machine, get a different number.”
Skip “AI is totally random” — a good program asked twice mostly holds steady, and that line is easy to fact-check. The point that holds: whoever picks the machine picks the number.
Want the receipts?
Everything above came from an experiment anyone can run on a laptop. Here it is, in three steps.