The biggest problem I've seen when people try to explain it is in the other direction, not people describing something generic that can be interpreted as a Markov chain, they're actually describing a Markov chain without realizing it. Literally "it predicts word-by-word using the most likely next word".
