From Philip Yancey’s Soul Survivor, the chapter on Annie Dillard:
I remember working on a short story early one morning. For three hours I strained to develop 3-dimensional characters and to purge all cliches from their dialogue. A raw beginner at fiction, I was getting a terrific headache from the effort. Naturally I used the excuse to stop writing and walk across the street to a coffee shop. Imagine my surprise when I discovered that all the people in the coffee shop were 2-dimensional characters who talked in cliches! None of them seemed nearly as interesting to me as the people who populated my story. I fled back to the security of the false reality awaiting me (and only me) in my basement office.
(The term 2-D and 3-D characters, BTW, probably comes from the earlier terms “flat” and “round”, which come from EM Forster’s Aspects of the Novel.)
This reminds me of a time many years ago when I was in the computer game AI business and wrote an article about computer AI characters. There was (and still is) a thing called the Loebner prize, which was supposed to advance AI by awarding prize money each year to the computer that does best at pretending, in a conversation though a terminal, to be a human.
Problem was that every year the winner was some stupid pattern-matching Eliza program that didn’t understand what it was saying at all. I said that stupid pattern-matching programs did better than logic engines at emulating humans because humans were more like pattern-matchers that responded to key phrases without really understanding what they were saying than like logic engines. I didn’t really believe it, but there’s an element of truth to it.
In fact I believe nearly the opposite: Everyone is interesting, if you know them well enough. But that’s hard to do. A novel presents a picture of its characters that isn’t as deep as real life, but is deeper than you get to see in real life.