The first thing I made with AI wasn't an app

I can't stop an AI from being confident and wrong about my work. I can stop it from being wrong about the facts, and that took a list, not a prompt.

A false number about my own work went out earlier this year. I went looking for the bug. There wasn't one: the model had copied the number correctly, out of my own notes. It was right in the file it came from and wrong in the three documents that quoted it, and it had no way to tell which one to believe. Neither did I.

Being more careful doesn't survive copy and paste. It wasn't judgment I got wrong, it was transcription, and it came back every time something got reused.

So I wrote the list. One file. Every number I'm allowed to say, sitting next to where it came from. Under that, the ones that are wrong and keep returning, written as the false version first and then the real one, because the false version is what my hands reach for. Then a short section of things that are true and still aren't mine to publish.

The check that reads the list came later and took an afternoon. The list took an evening, and it does most of the work.

If you've ever corrected an AI about your own field, you already know your first three entries.

What's the thing AI keeps getting wrong about your work?

Generated product reconstruction: the speaker and the screens are rendered, the files and the check are real. Synthetic narration.