Most of is like to think that a computer will never be able to write the greatest novel, surpassing War and Peace, Les Misérables, Ulysses, Remembrances of Things Past, and Middlemarch. A computer may long ago have beaten Gary Kasparov, the world’s leading player, at chess, and computers may be better able then ophthalmologists to diagnose macular degeneration; but a novel, a work of imagination, is a much subtler, more human thing. A “soul,” whatever it might be, is surely essential for writing a great novel. I fear, however, that a computer will be able to write the greatest novel.
The way that a computer will write the greatest novel is conceptually simple. We will ask readers to rank as many novels as possible. Indeed, we won’t need to ask them: we’ll simply to a website like Goodreads that has a score for all novels. But stop, I see a problem. Can we rely on this ranking?
Confusion by Elizabeth Jane Howard, which I’ve just read and hugely enjoyed, scores 4.24, while Middlemarch, surely a greater novel, scores 3.93. Indeed, Confusion scores more than War and Peace (4.1), Les Misérables (4.14), and much more than Ulysses (3.73); and one of the Harry Potter novels (none of which I’ve read, snob that I am) scores the highest of all (4.4).
So we need another means to rank novels. We’ll ask professors of literature from all the universities in the world to score novels. That will be a tremendous task, and inevitably English professors will outnumber Swahili and Gaelic professors. We can adjust for numbers, but we still have the problem that perhaps professors of literature are not the best people to rank novels: they are so weird and special. Then nobody will have read all the novels in every language. They will have a read a tiny and biased selection, and, as I’ve blogged before, things become famous not because they are the best. http://blogs.bmj.com/bmj/2016/11/21/richard-smith-why-is-the-mona-lisa-the-most-famous-painting-in-the-world-and-facebook-and-harry-potter-so-popular/
But let’s not let the best be the enemy of the good (although, Richard, we are talking about “the best,” so this is not very helpful.) Let’s start with novels in English only, including translated novels. Let’s start with one of those lists of the “best 100 novels in English.”
We then feed all these novels into a computer capable of “machine learning.” We ask the computer to identify the “rules” that make a good novel. We think that we know the “rules”: convincing characters, a compelling plot, “rich” language, accomplished story-telling, and deep insights. A computer can turn those “rules” into something can be acted on, in a way that we can’t. The computer will also identify “rules” that are not apparent to us. With these two achievements the computer has an advantage over the greatest novelist.
I’m on shaky ground here, as must be apparent, but it seems to me that we must feed the computer with both good and bad novels. If we could give them scores of rank them, then that would give the computer more data to process. And this I do know: the more data we give the computer the better—perhaps tens of thousands of novels. We might, for example, take buckets of Booker Prize winners, novels shortlisted, novels long listed, novels submitted, and all novels published that could be submitted that year.
The computer then might find it a fairly easy ask to choose next year’s winner, saving thousands of hours reading by the judges. But, Richard, isn’t originality, one of the essential features of a winner? Can a computer possibly identify originality from analysis of past novels? Isn’t that a logical fallacy? I’m not sure is my answer.
But even if the computer can identify the winner can it go further and “write” next year’s winner. We like to think, as humans surrounded by machines that are getting smarter and are poised to replace us, that that must be impossible. But is it? Is writing a novel fundamentally different from playing chess or just the same sort of problem only much bigger and more complex? Isn’t this a variant on the debate over whether consciousness is more than the interaction between billions of neurones? The latest thinking seems to be that the brain is just “a meat machine.”
But would a computer want to write a novel? Initially we would tell the computer to write a novel, but if we are replaced will it want to go on? We get pleasure (in its broadest sense, including lots of pain) from writing and reading novels; but what will be the motivation for computers? I speculated before on whether machines would want to go on once they had taken over humans or whether they would switch themselves off. As carriers of “the selfish gene” we are programmed to reproduce and keep going, but machines are not.
I heard Gary Kasparov on the radio this morning saying that our advantage over computers is that we have a “reason” (even if we are not sure or in agreement what it is) to carry on, whereas they don’t. So even if a computer could write the greatest novel it wouldn’t bother to do so.