The code AI shouldn't write

Today’s tools generate code at a speed that would have looked like science fiction a few years ago. They can write a whole CRUD, translate between languages, scaffold an entire project in seconds. The interesting question is no longer what AI can write, but what it’s wise to let it write.

It isn’t an ideological distinction. It’s practical.

Where judgment isn’t delegated

Some code has an asymmetric cost of being wrong: cheap if it works, ruinous if it fails in silence. The logic that decides who can access what. Handling money. Migrations that touch data you can’t recover. The boundaries where your system trusts something external. In all these places the bottleneck was never writing the code — it was understanding exactly what it had to do and why.

AI speeds up the writing part. The understanding part is still yours, and it’s the expensive one.

Generating code fast and understanding it well are two different jobs. Confusing them is the mistake.

Accepting isn’t the same as generating

The healthy shift isn’t to ban the tool, but to move the weight of the work. Let AI propose; let the human decide whether that proposal is the right one for this system, this domain, this team. Accepting a diff demands the same judgment it would have taken to write it — only now it’s easy to skip that step, because the code is already there, plausible and tidy.

The speed is real and worth it. But direction is still a human decision, and delegating it is the one mistake the machine can’t help you fix.