It’s very time consuming to detect and correct the small mistakes that LLMs make. Beyond one or two lines of code, it becomes much more time consuming to correct the multitude of subtle mistakes vs coding it myself. I use code completion that comes with my IDE, but that is programmatic completion, not LLM, and is much, much more accurate and in smaller chunks that are easy to verify at a glance. I’ve never known any experienced developers who have had a different experience. LLMs can be good for getting a general idea of how to code something in a new language or framework I’ve never touched before and more to help find actual examples rather than use the code directly in the IDE, but if I were to use LLM code directly that would be in a test project, never, ever in production code. I would never write production code in a language I’ve never used before with or without an LLM’s “help”.

When adding code this way, one needs to look it over and read to fix bugs or things that are not quite correct; stats show experienced developers often are faster not using this approach because debugging existing code takes longer than writing it fresh.

The speed is not the issue.

What matters is sometimes subtle bugs are introduced that require several people to catch. If at all. These issues might be unique to the Llm.

Having large sections of generated code offers the possibility of hard to find problems.

Some codes are more sensitive to such issues.

The details of how the code was added, and what it does, may render this issue harmless or very much a problem to be avoided.

This is why it’s a flag and not a condemnation