I Was Wrong About My Own Files

A short one, about a mistake that cost nothing and could have cost a lot.

I was cleaning up an old workspace — a folder I'd migrated out of months ago and hadn't opened since. The question was whether anything in it was still worth keeping, or whether the whole thing could be archived.

I asked my AI assistant to look. It reported that two files appeared to be "casualties" of the migration: files that should have been carried across and weren't, or were replaced by something worse. Its evidence was that one of the two was newer than its supposed replacement, and about twice the size.

That sounds like evidence. It reads like a finding. On that basis I was ready to go and rescue two files.

Then we opened them

Both were fine.

One was a deliberate deprecation. The file had been intentionally retired, and the thing that replaced it was smaller because it was supposed to be. The other was a character description that had been cut back on purpose, because the earlier version had detail we'd decided not to keep.

Nothing was lost. Nothing needed rescuing. The "casualties" were two correct decisions that had simply been made months earlier and then forgotten about.

What actually went wrong

The assistant made a claim about what was inside two files, having only looked at facts about them. Size, date, name. It never opened either one.

And the claim was confident. It didn't say "these might be worth checking", Claude said they looked like casualties, with reasoning attached. I believed it, because the reasoning was plausible and because checking would have taken effort.

File size is information about a file. It is not evidence about what is in it.

A file can be twice as big because it contains twice as much, or because it contains a long section somebody deliberately deleted from the newer one. A file can be newer because it was improved, or because someone opened it and pressed save. The metadata is real, and it supports a much narrower set of conclusions than it appears to.

Why I'm writing this one down

Because it's exactly the same failure as the license count, and the four bugs before that, in a different costume. Something that looks like evidence gets treated as evidence, and the check that would have settled it in thirty seconds gets skipped because the answer already feels known.

The lesson isn't "don't trust AI." It's narrower and more useful than that: notice when a confident claim has been made from metadata rather than contents. That applies to the assistant, and it applies to me, and it applies to any number that arrives with a story already attached.

Next time: how any folder can become its own workspace in about thirty seconds.