Last week, two researchers announced that they had decoded the science of writing.
More specifically, Jodie Archer and Matthew L Jockers had used large datasets to determine which features were most common in bestselling books. The pair insist their information can’t be used to reverse-engineer bestsellers. But if scientists can use code to break down novels, it stands to reason that they can’t be far off writing the perfect novel; one most likely to achieve astronomical sales, and make those scientists rich.
At first, a computer-generated book might sound like the latest instance of the creeping hand of the automaton. But I see this a little differently. I think this story gives book-lovers enormous cause for optimism.
No robo-books – for now
Scientists will have to clear so many hurdles before they develop an artificial intelligence capable of writing a novel.
Let’s leave aside the obvious question of motivation. Scientists tend to be quite clever people – probably too clever to turn to print journalism as a source of wealth. This research might make for good headlines, but it won’t much help the bottom lines of cash-strapped university labs.
Then again, scientists do weird stuff all the time. Who’s to say a team, somewhere, with nothing better to do, won’t attempt to use AI to write a novel?
At present, AIs are indeed capable of writing at the same level as humans – at least in some fields. Financial news – especially on topics like share price movements – can be produced near-instantly, and word-perfect, by algorithms.
That makes total sense. An editor at Thomson Reuters or the Associated Press could teach anyone at all to do that job in 30 minutes. Training an AI to switch certain terms (company name, ticker, percentage price change, new stock price) into a template is not difficult.
Archer and Jockers’ research works on the assumption that novels can be boiled down to similar templates. Now, some types of literature – like John Grisham thrillers or “bonkbusters” – may be formulaic, but recreating them at every level of language is a serious ask.
It’s quite possible that algorithms and formulae exist that can quantify the abstract concepts upon which novels depend. I’m sure, for example, an algorithm could be devised that churns out perfect examples of suspense by analysing the number of words that pass, in thrillers, between a murderer entering a room unseen and a victim being killed. That – though impressive – is only a tiny part of the novelist’s job.
Let’s say an AI succeeded in deploying all the writer’s formal strategies, and produced a book that chose the right words in the right order, put them in sentences that obeyed grammatical rules, and combined them to form a plausible narrative. A tall ask, but possible. Unless the AI behind the program was a genuine, broad AI, it would be incapable of producing original work. It could combine structures and plots that already exist in books it had learned – but it could not produce original work.
As somebody who loves words, I’ve got a bit of skin in this game. So, let me use another medium to show just how far away proper robot literature is.
The “perfect pop song” is total garbage
Every few months, scientists will claim to have reached the pinnacle of culture in one form or another. Take this music.
DarwinTunes evolves music over thousands of computerised generations. Over time, the music gets better, more complex, and more pleasing to the human ear. After one generation, it’s rough. After a hundred, a bit better. And after 10,000?
Still pretty bad. The track sounds exactly as you might have expected: like a computer wrote it with no instructions beyond replication of the sort of things we like. It’s music, but it’s not very good. I certainly couldn’t listen to it for an extended period of time.
So, why can’t computers write music or novels yet?
Maths eats everything – and it’ll eat writing, too
If you were asked to name something a computer couldn’t do, you probably wouldn’t plump for “doing equations”. Even a mediocre computer would eat the world’s finest arithmetician alive. If you had to beat a computer at something to save your life, you’d probably choose something creative, or abstract. Something that couldn’t be quantified.
It’s not a coincidence that at any school or university, arts students go to libraries, while science students go to computer labs. Over the course of the information age, science has increasingly come to mean that which can be quantified, measured, and taxonomised.
That’s exactly why mathematics is considered “hard” science. So is physics. Chemistry is a science. Biology? Sure, but you wouldn’t exactly brag about your child taking it for A-Level.
What about the social sciences?
You’d probably agree that psychology or sociology is less demonstrably scientific than algebra. But how does the picture change when we apply powerful computational methods to those areas? It’s hard to argue that an AI which can learn to spot depression by mining datasets for almost 200 variables isn’t scientific.
Similarly, anyone who follows Nate Silver’s excellent FiveThirtyEight blog – which predicts elections at all levels with startling accuracy – will know that sociology, when combined with computers, can be rather more clear (and thus, more obviously useful to the average person).
Historically, an area of inquiry has become a science the closer it has come to mathematics. Naturally, an increase in computing power has made this process faster. That’s why analysis, like that of Jodie Archer and Matthew L Jockers, is possible. And though they haven’t yet mastered the art of literary production, Archer and Jockers are two of many scientists working to reduce (or elevate) literature to a scheme of ones and zeros.
For now, books and mathematics are largely considered to belong to two separate domains. But just as maths has eaten every other field, it’ll damn well eat literature too. Whether or not a computer can truly read or write books may seem irrelevant; but in reality, it might be one of the best measures we have for recording true computer intelligence. Once a machine can analyse art like it analyses numbers – when it sees the entire world in vivid binary – then, and only then, will it be truly intelligent.
Category: Artificial intelligence