AI demos and tutorials almost always start from a fresh, clean codebase. "Look how easy it is to build a ToDo app with AI!" Real codebases rarely look like that. They are old systems built over years by people who have since left, with naming that changed halfway through and documentation that barely exists.
That is the brownfield reality, and there the AI behaves like a confused intern. It suggests refactors that break six layers of dependency injection. It tries to modernize code that is load-bearing for reasons no one remembers. It doesn't know the *history*.
The habit that worries me is pasting the error into the chat and accepting the first answer, treating the model like an oracle. The code mostly works, so nobody asks why, and debugging a race condition stays a skill nobody had to learn.
We are also flooding the internet with generated noise and wrecking its signal-to-noise ratio ourselves. Choosing velocity over quality only piles up technical debt.
So yes, use the tools. I use them. Just stop trusting them blindly. Learn the fundamentals, including the memory model, and read the full docs. When the model hallucinates, and it will, *you* are the one who has to fix it, and you can only fix what you understand.