What stops captchabots from just brute-forcing it with common words? Or, if they wanted to get fancy, something like a Markov model of English text could really help narrow down the likely missing words.