The Bug I Couldn’t Blame on the Model

I shipped a bug the AI wrote.

I’d asked for a function, it produced one,
it looked right,
and I dropped it in without really reading it.
Days later it failed in production,
in a way that was obvious
the moment I finally looked.

My first instinct was almost funny.
Some part of me wanted to say,
“well, the AI wrote it.”

And then I heard how hollow that was.

The AI didn’t ship it. I did.
The AI didn’t skip the review. I did.
My name was on the pull request,
not the model’s.
And the model wasn’t getting the alert at 2am.
I was.

That was the moment
I stopped treating AI output
as something I could hide behind.

Now I treat everything it gives me
like a draft from an eager junior —
fast, confident,
and sometimes wrong
in ways that look completely correct.

I read it.
I question it.
I test it like a stranger wrote it,
because in a real sense,
a stranger did.

The value was never
in it generating the code.
It’s in me being the person
who can tell
good output from merely plausible output.

The tool has no stake in whether it’s right.

I do.

And that part
was never going to be automatable.

– Serguey Asael Shinder

Leave a comment