"Ironically" because I usually rewrite almost everything it generates (sans the file and class names, sometimes). But when I describe the idea enough to give Claude some chance of implementing it, I refine it and start planning the high-level implementation. Then the file of LLM slop, even though I have to rewrite most of it, is much less overwhelming than a blank file.
Despite this, AIs get fooled to this day. There are still jailbreaks for GPT-5 and nudity and piracy on YouTube.
The only way to distinguish “good” from “bad” is competence, which has never existed on a large scale.