You can’t practically “trust but verify” with LLMs. I task an LLM to summarize an article. If I want to check its work, I have to go and read that whole article myself. The checking takes as much time as just writing the summary myself. And this is even worse with code, as you have to be able to deconstruct the AI’s code and figure out its internal logic. And by the time you’ve done that, it’s easier to just make the code yourself.

It’s not that you can’t verify the work of AI. It’s that if you do, you might as well just create the thing yourself.