There have been quite a few system outages lately -- Microsoft, the FAA, the banks, Apple -- and a lot of speculation about the cause. To those of us in the industry who are a bit more (ahem) senior, it’s pretty obvious.
Nearly all of the outages have happened as a result of system “upgrades”, and most of those upgrades certainly included sloppy AI-generated code.
AI has changed the way we write code. (I’ve been doing a lot of that lately, now that my political campaign is over.) AI generates enormous volumes of utter garbage code, misleading advice, and fragile, bug-ridden dreck.
(Yes, and it needs to get off my lawn.)
This week I’ve been rewriting the code that manages the internal data structures in my latest project. I used AI to analyze the code, and it found some issues around locking and durability in a heavy-traffic context. Okay, terrific, thank you for telling me, we need to fix that.
Normally, without AI, I would just trace through the code myself, understand it, write a few tests to demonstrate the problem, and come up with some minimal and hopefully elegant fix.
With AI, though, we don’t do it that way. We say to Claude or Codex: go ahead and fix the problem. And it does, by generating very, very large amounts of overly complex code. The code sort of does the job, but it misses the creative fix that would have left us with a clean, simple result. In the industry, we call this technical debt: it’s like borrowing money to get what we want now, at the cost of having to pay big penalties later.
AI is seductive, really seductive. It fixes the problem in minutes. And if you close your eyes and just commit the code to production, you can tell yourself you’re done, and you don’t have to understand the logic behind what it did. It’s like using Google Maps every time you drive. You don’t have to know where you are or how to get home. You have no mental model of the map.
So I used AI on my data structures problem, and I spent three days this week trying to untangle what it did. It told me that my proposed locking fix wouldn’t work, and that I had to implement this very long, convoluted algorithm to avoid an unrecoverable crash. So I argued with it, and over and over it apologized for having misled me, and then proceeded to get it wrong again. I cursed at it, sometimes in all caps, because I learned that, bizarrely, cursing actually yields better results. It makes it more likely that the AI will actually go back and read the instructions I gave it.
Cursing at inanimate objects comes naturally to me. My dad fixed cars for a living, and I would frequently hear him cursing into the wind whenever I visited the garage. He got angry, over and over, at poorly designed cars where simple jobs took hours because some Detroit engineer with clean fingernails made it impossible to get at the part that needed to be replaced. I’ve pretty much turned into my dad.
AI sucks, but you can’t avoid it in this industry. The analogy I’ve been using lately is that an AI coding assistant is like having a free moron sitting next to you all day. Sometimes he gets it right. He types fast. He’s enthusiastic. And he costs almost nothing, so you can’t not use him.
But if you use him and you don’t know what you’re doing, he generates the kind of slop that will take down a network.
I am certain that that is what happened this week, at least with some of these outages. Engineers are doing upgrades with sloppy, untested code.
The only question is whether management at these companies will stop making them use free morons to do the work.

