But with a compiler you can see why the compiler did what it did, with AI you have a black box. Even if you fixed the AI model to be deterministic you could still not see why it produced the specific output for that specific input.
The point is that it's reasonable and possible to predict the compiler's assembly output from high-level source. Maybe each individual can't do it, but they could, with a little learning.
With LLMs, you just can't predict the output for a given prompt, and can't predict how changing that prompt will change the output (or how much it will change the output). Even the people who build and work on LLMs every single day and know them intimately can't.