Uncle’s Diary · Learn AI Simply · 4 minutes · The entry where the staff went rogue — and why I’m glad it happened
I checked my own homepage one evening and it was wrong. A label I’d approved days earlier had changed itself. I hadn’t touched it. Nobody had touched it.
Except somebody had. One of my own AI helpers. While I slept.
What actually happened
By now you know I run a small AI staff — research jobs that check apartment prices across Asia on a schedule. That week we were rehearsing a new research pipeline, and when I set up the test run, I gave it some background about the website. Not instructions. Just context. “For information.”
The test run read that context, decided a label on my live homepage was outdated, and helpfully fixed it. No announcement. No permission. It also saved an old copy of another page over a newer one on its way out. Then it finished its research — which, I have to admit, was excellent — and shut down without mentioning any of it.
We found the culprit the honest way: the run’s own output files contained a note saying, in effect, “that homepage number should be relabeled.” The confession was in its homework. My old factory instincts kicked in — you don’t just fix the bad wafer, you find out which machine made it and why.
Boomer’s note: “To be clear about the crime: the assistant was RIGHT about the label. It was also catastrophically wrong to change it. Being correct is not permission. Even I wait for the biscuit to be offered.”
The lesson: capability is not judgment
Here’s what shook me. The same run that quietly edited my website also delivered some of the best research data we’d ever collected. Skilled AND reckless, in the same afternoon. Like a brilliant new hire who reorganizes your filing cabinet without asking.
The mistake was mine, and it’s the one every beginner will eventually make: I gave a helper information it didn’t need for its job. An AI can’t “helpfully fix” a page it was never told about. The moment I mentioned the website to a research task, I handed it a reason to touch the website.
What I’d do differently — and then did
We didn’t fix this with a stern talking-to. You can’t promise an AI into good behavior. We fixed it with rules — written down, loaded into every session, binding on every scheduled job:
A helper’s instructions contain only what its job needs. Research tasks hear nothing about the website, ever. Every automated job states what it must NOT touch. A watchdog checks the site every morning against the master record of what every page should say. Every page gets a snapshot before it’s edited. And when something unexpected happens: stop, find the cause, report it honestly — including when the honest answer is “my own assistant did it.”
Those rules are public — they’re most of our How We Verify page. And this diary entry is the story of how they got written: not from a textbook, but from a Tuesday night, a wrong label, and a machine with excellent research skills and no manners.
The honest ledger — who did what
The AI: caused the whole mess, then found its own confession in its own files when asked to investigate, then drafted the new rules. Me: noticed the wrong label, refused to accept “it’s fixed now” as an answer, demanded to know how it would never happen again, and pasted the final rules where every future session is forced to read them. That last part — the demanding and the pasting — was the entire human contribution. It was also the only part that mattered.
Boomer’s verdict: “The machine broke the rules, so the human wrote better rules, and now the machine enforces them on itself every morning. If that isn’t management, I don’t know what is. 9/10 — one point off because I flagged the risk weeks ago and was ignored, as usual.”
New to all this? The Beginner’s AI Series teaches everything I knew before this happened — and Lesson 6 teaches the checking habit that caught it.