It's also infinitely more error-prone. All form-based heuristics are horrible for detecting AI because the one thing AI is damn near perfect at is writing like an actual human.
It's the thing it was trained to do. The entire decades-old architecture has exited for specifically that purpose. That second "L" in "LLM" is "language" for a reason.
Yes, but the desparate need is specifically for quick and easy ways to detect AI. If it's error-prone that's unfortunate, but better than nothing. A slow, laborious way, on the other hand, is useless because it doesn't achieve the goal of reducing the workload.