My take on using AI to code

I use AI to write code every day. It makes me faster, mostly on the parts I already know how to do and where I do not feel like typing them one more time. That part is real and I am not going to argue against it.

And after that everybody says the same sentence: you have to understand what it wrote. Yes. Of course. I never met a developer who says the contrary out loud. It is like saying you should test before you push. Everybody agrees, nobody convinces anybody, and the discussion stops there.

For me it stops one step too early.

When you push code in production, it is your code. Not the code of the model. If something breaks, a small thing or a big thing, it is you who broke it. Nobody is going to open the incident and write that the assistant did it. You can try one time. You will not try twice.

So this is where I put my line. If it gives me something I cannot read, or something I could not have written myself with more time, I do not keep it. Maybe it is even better than what I would have done, tant pis. I cannot answer for it, and someone has to answer for it.

The important word here is time. I am not saying I must write it as fast, I am clearly not. I am saying that with a week and the documentation, I should arrive somewhere near. If I cannot, then it is not really my code and I am only the person who pressed enter.

I am not against AI, I would not open it every day if I was. But I do not put in production a thing I cannot defend. If I cannot explain it in a review, or repair it at three in the morning when it wakes me up, it should not be there.