Yeah, this tracks with what I've seen. The problem isn't the AI—it's that these tools optimize for "looks good in preview" rather than "works in production." They nail static layouts but choke on state management, error handling, and actual data flow.
I've had better luck using Claude or ChatGPT as a code review partner than as a generator. Feed it your architecture, get it to spot issues. Generated boilerplate tends to work only when you're okay with throwing it away and rewriting the non-trivial parts anyway.