I think it's mostly because the output is often confidently incorrect. So we need to spend effort fact checking each time we use it, which defeats the purpose of asking the machine in the first place.