People need to "keep AI on the leash" because current large language models still make mistakes no human ever would, said ...