The media is full of hype about technology in general, and artificial intelligence (AI) in particular. The Robots are Coming for our Jobs scream the headlines, and tech luminaries like Elon Musk warn us that super-intelligent computers could exterminate the human race. In this febrile atmosphere, it seems that no domain is safe from the incursions of AI, as proved by the recent New York Times article about a writer who uses AI to help finish his sentences. This does not, on the surface, seem like a particularly noble endeavour - my wife and I regularly finish each other's sentences without needing expensive algorithms to help us. But it's indicative of a wider trend, as companies seek to automate the production of text without pesky humans being involved.
But what's the reality of current efforts to write using computers? And will they eventually supplant our own efforts?
In order to answer these questions, it's necessary for me to take a short detour by explaining what AI actually is. Don't worry, I'll try to keep this understandable for humans by including cute animal pictures!
Photo by Smerikal |
What we currently call AI is actually a technique called Machine Learning (ML). There are a few types of ML, but the version most used is called Supervised ML. To understand how it works, imagine a guide dog. It starts out life as a puppy - cute but undisciplined. A human trains the puppy to follow basic instructions, walk in a straight line and react to the dangers that exist in the modern world. But the learning doesn't stop there. Once the guide dog is given to its owner, it will have to constantly appraise unfamiliar situations and hazards, and react appropriately.
The ML algorithm is like the puppy. Well, except it's a computer program, of course. The first stage of supervised ML is the training phase. A small set of training data is fed into the system, and the algorithm creates a certain type of output from it. A human will then assess that output, tweak the algorithm and run the process again. Once the human is satisfied (which can often take a long time) the algorithm is fed with real-world data (the more the better). Although the algorithm has never seen this real-world data before, it can make choices based on what it has learned during the training phase and create a completely new output from it.
Seen from this angle, ML isn't actually that clever. It relies on humans to write the algorithm, and supply the output format and training data. But the technology has been hyped to dangerous levels, as this Guardian article explains. The real strength of ML is that it can make decisions based on vast amounts of data that would take a human a lifetime to digest. If you subscribe to the theory that a writer is just the sum of their influences, the idea is that you could feed loads and loads of existing works into an algorithm and have a new one pop out the other side. In practice, it's rather more difficult than that...
In 2016, director Oscar Sharp and AI researcher Ross Goodwin set out to make a short film written entirely by a computer. Called Sunspring, it was created by an ML algorithm called Benjamin. The training data set was a series of prompts from a sci-fi filmmaking contest. The input data set was hundreds of sci-fi movie screenplays. And the output was a movie script, which was then staged and acted by professional filmmakers. Here's the result:
Well, that was "special". Not so much a script in fact, more a collection of lines of dialogue and action cut-and-pasted together. My favourite quote is:
"But I'm the one who got on the rock with the other two, and left a child."
I also love the fact that the film features an actual Chekhov's Gun, duct-taped to the wall. It's absolutely hilarious watching the actors doing their best to emote with a screenplay that is borderline gibberish, and it makes you realise how much an actor's performance brings to a movie.
OK, not so impressive a performance from our AI screenwriter that time. In June of this year though, the team behind Sunspring tried again, with the twist that this time they gave full control to the AI. As well as loading up the AI with movie scripts, they gave it green-screen footage of the Sunspring actors and actual public-domain movies and music. This is Zone Out, the short film that emerged from the process:
This one is much more interesting, possibly due to how bizarre it is. The results are a lot like watching a David Lynch film (Eraserhead springs to mind) but with freaky face-swap technology mixed in. Zone Out is genuinely unnerving in places, though it derives a lot of its power from the mise en scène of the original movie footage used (particularly The Brain That Wouldn't Die). However, one thing that has remained constant between the two films is the quality of the script - it is woeful!
On the evidence of these films and other experiments, ML has a long way to go when it comes to writing fiction. Meanwhile though, AI has been gradually creeping into journalism. Obviously, an algorithm can't write an opinion piece yet, but they are very good at cranking out copy based around predictable subjects. For instance, content generation firm AI Insights has this case study about their work with Yahoo Sports, claiming that 70 million sports reports and match recaps have been created using their technology. You might expect such writing to be bland, but AI Insights have given their ML algorithm a distinctively sarcastic voice, which helps to mask the fact that the content has been generated by a computer.
As we all know, technology moves quickly, and it's hard to be sure how it might develop. But what are some likely next steps for machine writing? This Deadline article by Arvin Patel has some fascinating but grounded ideas about how AI might affect Hollywood. A lot of them aren't about replicating tasks we already do, but creating new forms of content, like TV series that are uniquely created for your own interests (this idea of content personalisation can also be seen in the Yahoo Sports article).
Could we one day have a novel that rewrites itself to suit the reader's likes and dislikes? I actually imagined this scenario way back in 2013 for a short story called Mindworm, which you can read for free on my website. Luckily, nothing like this has happened for real, as yet...
Mindworm illustration by Mei-Li Nieuwland |
I think a much more likely scenario for machine writing is the creation of new works by dead authors. You can imagine a situation where all of Jane Austen's novels, letters and half-finished manuscripts are fed into an ML algorithm to create an entirely new book in her authentic voice. Or how about a "previously undiscovered" Shakespeare play? The publishing industry have been churning out this kind of thing for years using ghost writers, so the idea they might do it with algorithms isn't too far-fetched.
As for the technology supplanting us living fiction writers, I reckon we can breathe easy for now. This is because a writer isn't just the sum of their influences - we absorb the content and then apply our own unique perspective to it. That perspective is formed by a cocktail of experience, consciousness and emotion that is currently impossible to synthesise. Existing ML algorithms can only remix existing content, they can't create something wholly new. To do that will require a computer that's able to think entirely like a human, and that technology is as far off as it ever was.
Nick.
Nick Cross is a children's writer/illustrator and Undiscovered Voices winner. He received a SCBWI Magazine Merit Award, for his short story The Last Typewriter.
Nick is also the Blog Network Editor for SCBWI Words & Pictures magazine. His Blog Break column appears fortnightly on W&P.