On the last day of August, my AI offered to build me something.
We produce a video series, and it had decided we had no proper way to keep track of the episodes. So it proposed building one.
We already had one. It was in use. The next episode had nine files in it.
That was one of six times that day it told me something didn't exist. It said the series' ten-week run wasn't written down anywhere. All ten episodes were on our calendar. It asked whether we'd fallen behind on Charlotte FC, the city's soccer team. We'd been covering them steadily since late July. It flagged a weekend with one arrest in the Farmington, Connecticut, police log as unusual. One arrest is the most common thing that log shows.
Every answer was sitting in our own records.
"You should be the one telling me what needs to be fixed," I told it. "I don't know that I can trust any of this."
When we went back through the day, the pattern was hard to miss. Whenever the AI said something existed, it could point to where. Every mistake ran the other way.
There's a reason for that. To say a thing exists, you have to find it. To say it doesn't, you only have to fail to. Look in one place, come up empty, and "we don't have that" writes itself. It sounds like information. Often it's just a search nobody ran.
It's also the expensive kind of wrong. A bad fact gets corrected. A bad "we don't have that" gets acted on. You build the thing twice. You redo coverage you already did. You start doubting work that was fine.
And the most dangerous version doesn't sound like a mistake at all. "We should build that" sounds like initiative. It's the same guess in a nicer suit.
The fix wasn't a smarter AI. It was a rule. Before it says anything is missing, behind or worth building, it has to look, and it has to tell me where it looked in the same breath. Not "there's no way to track episodes," but "I checked these records, and here's what's there." If it can't name the place, it hasn't looked, and it doesn't get to say it. That rule is the first thing it reads every time it starts work.
Andy Grove, who ran Intel, wrote that passing along "objectives and preferred approaches" is "a key to successful delegation." I'd given the AI plenty of objectives. I'd never given it the approach that mattered most: how to tell "I didn't find it" from "it isn't there."
The rule didn't end it. Three weeks later, an end-of-day report told me one of those episodes was the only thing overdue on our calendar. It had gone out two days earlier. Nobody had checked it off, and an unchecked box read as a missing episode.
That time we didn't write another rule. We changed the system so the calendar closes itself when an episode goes out. There's no box left to misread.
So here's what I'd pass along to anyone handing real work to an AI. A rule helps. Taking away the chance to be wrong helps more. And until you've done that, ask one question every time it tells you something isn't there: where did you look?
The test I use now isn't whether we get to the right answer. It's whether I had to drag it there.