Totally guessing here but this would be a property that gradually appears as the model size increases.

With a tiny model you would get gibberish and with ever increasing models the response would increasingly approach a coherent answer to a finally correct answer.