Mistakes AI made
If you only read the enthusiastic posts about AI, you'd think it never gets anything wrong. It does. Over two weeks of real work on my network, Claude made a handful of mistakes. None of them did any damage, because we caught them all. But how they got caught is the interesting part.
The list
| What it said | What was true | Who caught it |
|---|---|---|
| A backup key could be restricted so it can add backups but never delete them. | The backup tool needs to delete its own small lock files, so that kind of key doesn't work. The right answer was a storage-side lock that keeps everything for 30 days. | Claude, before I'd set anything up |
| The first big backup would take two and a half to three hours. | It took an hour and seventeen minutes. Much less data was new than it assumed. | The backup itself |
| If a backup was still running when the next one was due, the next one would be skipped. | It was queued instead, and ran straight afterwards. | Claude, from the logs |
| Older backups of my Windows machine would be kept for months. | Under our retention rules, most of them would be dropped after a week. | Another Claude session, working on the Windows side |
| A leftover lock from an interrupted backup would be ignored automatically. | Backups ignore it; the cleanup step doesn't. | The next morning's error email |
| An unknown device on my network was probably a development board. | It was a smart indoor herb garden. | Claude, once it asked the device itself |
| "I'll finish the router switch once the old router is off." | It runs on a computer that reaches the internet through that router. When the router went off, so did Claude. (The whole story.) | Me, when it went silent |
| My laptop has a USB port on the right side. | It doesn't. All its ports are on the left. | Me |
| Click "Upgrade" to install the router firmware. | The button says "Update." | Me |
| "Don't put the new router on a switch." | It meant only the modem side. The house side should go through a switch. The wording was just unclear. | Me, by asking "why?" |
And one near miss of my own: when I went to download firmware for my Wi-Fi access point, the file I found looked right but I wasn't sure, so I asked. Claude had me check the hardware version on the device itself, and the file turned out to be for a different hardware version. We found the right one before anything was installed.
What I take from this
The mistakes sounded exactly like the correct answers. That's the real lesson. There was no hesitation in the voice when it was wrong. If you can't tell from the tone, you have to tell from the evidence.
Most of them were caught by checking, not by me being clever. A log, a test, an error email, a second look. The work was set up so that claims got checked against reality, and that's what caught the errors. Building in those checks is worth more than trusting anyone's confidence, mine included.
It admitted mistakes plainly. Every time, Claude said "I was wrong about that," explained what was actually true, and fixed the notes. It didn't hedge or bury it. I've worked with people who found that harder.
Ask "why?" Several of these surfaced because I asked a simple question about an instruction that didn't make sense to me. That's an old engineering habit, and it works just as well on a machine.
None of this makes me trust AI less. It makes me trust it the way I'd trust a very good colleague: a lot, with verification.