I’ve seen a lot of comparisons likening LLM code generation to how compilers and high level languages eventually removed the need for programmers to understand machine code. Arguments about determinism and code quality aside, I still think it’s a false equivalence. Even if LLMs produced beautiful, ideal code, they fundamentally remove a programmer’s understanding of what exactly is present in the logic. Mishaps are inevitable.

To prevent them, experienced programmers will do what it takes to produce human understanding and quality in spite of code generation. Yes, read the generated code. Yes, do code reviews with trusted peers. Yes, write/generate unit tests. There are even novel ideas about having LLMs interview us to make sure we understand the generated code’s purpose. That’s not the same as compilers, and it’s not the story AI companies have told us.

Just like in the beginning: Programming languages are made for programmers, not for machines. It’s on us to understand.