If Sunspring is anything to go by, artificial intelligence in film-making has some way to go. This short film, made as an entry to Sci-Fi London’s 48-hour film-making competition in 2016, was written entirely by an AI. The director, Oscar Sharp, fed a few hundred sci-fi screenplays into a long short-term memory recurrent neural network (the type of software behind predictive text in a smartphone), then told it to write its own. The result was almost, but not quite, incoherent nonsense, riddled with cryptic nonsequiturs, bizarre turns of phrase and unfathomable stage directions such as “he is standing in the stars and sitting on the floor”. All of which Sharp and his actors filmed with sincere commitment.
“In a future with mass unemployment, young people are forced to sell blood,” says a man in a shiny gold jacket. “You should see the boy and shut up. I was the one who was going to be a hundred years old,” replies a woman fiddling with some electronics. The man vomits up an eyeball. A second man says: “Well, I have to go to the skull.” And so forth. An unwitting viewer might be unsure whether they were watching meaningless nonsense or a lost Tarkovsky script.
This isn’t how artificial intelligence is changing the movies – yet. But AI is changing the industry in other ways, to a greater extent than is being admitted. In early January, Warner Bros broke cover and announced it had signed up to an AI-driven project management system that would “inform decision-making around content and talent valuation to support release strategies”. In other words, Warners will be using AI to help it decide what movies to make.
The system in question was launched last year by Cinelytic, a Los Angeles-based startup. Other clients listed on its website include Sony Pictures, Ingenious Media and STX Entertainment (the studio behind Hustlers and Playmobil: The Movie). Meanwhile, 20th Century Fox has partnered with Google to develop its own AI, named Merlin, which predicts audiences based on analysis of patterns and objects in movie trailers. Other new companies have also entered the AI game in the past few years, including Belgium-based ScriptBook and Israel’s Vault AI.
You can see the appeal, especially for a studio like Warner Bros. As well as putting out hits such as Joker and A Star Is Born, it has had its share of box-office flops, such as King Arthur: Legend of the Sword, The Kitchen, Shaft and The Goldfinch. These projects must have looked like hits on paper. How to tell the difference? Traditionally, Hollywood almost prides itself on its own unpredictability, accepting as a given William Goldman’s adage “no one knows anything”. It is an industry that carries out extensive analysis, research and audience testing, but that still relies to a large extent on human intuition and gut instinct. It is also an industry where the majority of films made are never released, and less than half of those that are make their money back.
Can the machines do better? Absolutely, says Cinelytic, which claims an accuracy of 85% in its box office forecasting. Cinelytic was founded in 2015 by Tobias Queisser, whose background is in finance and film producing. His co-founder, Dev Sen, is a rocket scientist who developed risk assessment software for Nasa. Cinelytic claims to have crunched data from more than 95,000 movies and 500,000 actors and professionals, but its chief selling point is that it can make forecasts in real time – expressed as percentage probabilities of certain levels of success. With its user-friendly interface, clients can play with the variables and assess the impact on box office right away.
Let’s make an action comedy starring Dwayne Johnson, say. It will be huge! But what if Johnson is not available? How would it look if you cast Gerard Butler instead? Not so huge, perhaps, but the budget might also be lower. How would it do if you released it over 1,000 screens, or 3,000? How would it do in Brazil or China? In effect, Cinelytic’s AI treats the movie industry like fantasy football, assigning quantitative scores to individuals, according to factors such as recent or past box-office performance or social media profile.
Others take a slightly different approach. ScriptBook’s AI primarily analyses scripts, as opposed to actors or directors. “Most people believe that cast is everything, but we’ve learned that the story has the highest predictive value,” says Nadira Azermai, who founded the company in 2015. It doesn’t take a computer-sized brain to think of a movie that had a stellar cast but still flopped. But a good story will succeed even without stars – and potentially more so with them, she says.
Within six minutes, ScriptBook’s AI reads a script and assigns it more than 400 parameters. “Anything that is information: emotion analysis, the journey for the protagonist and antagonist, whether the film will cater to a wide audience or niche audience, whether it follows a traditional three-act structure, whether the action happens in the most important places.” As more information becomes available, such as the cast, the prediction becomes more accurate, but if the computer says no to begin with, no additional information will change it to a yes.
ScriptBook’s success rates are comparable to those of Cinelytic: 83% to 86%, Azermai claims, whereas human decision-making is successful 27% to 31% of the time. In a retroactive study of Sony’s output between 2015 and 2017, for example, it successfully identified 22 out of its 32 releases that lost money. It could have saved the company a fortune. It has achieved similar levels of accuracy with movies before they were released, even predicting the success of unconventional outliers such as Get Out, La La Land and A Quiet Place. Azermai is not yet at liberty to reveal which companies ScriptBook has been working with, but she applauds Warners’ announcement that it is working with Cinelytic. “It is an important message, because the majority of companies are still treating AI as a dirty little secret.”
One of the factors driving the AI revolution is undoubtedly Netflix, which, by its spectacular ascent over the past decade, has awakened the entertainment industry to the awesome power of data. Netflix has valued its “recommendations” algorithms (which suggest to viewers what to watch next) at $1bn (£760m) a year in terms of keeping subscribers engaged. The company is notoriously secretive about its methods, but it is safe to assume it is tracking viewers’ every move: what they watch, how long they watch it, what time of day it is, what they watch next, how many hits a new show receives in its first day, and so on. Because Netflix also streams other people’s content, it is even better informed when it comes to producing its own. The studio’s staggering output (Netflix put out 700 original TV shows and 80 features in 2018) would not be possible without AI assistance. As a result of Netflix’s success, tech companies such as Apple and Amazon are getting into content creation and giving the old-school studios a run for their money.
“All of a sudden you have this conservative, traditional industry versus companies who are believers in data,” says Azermai. “There is a war on content, but one party is using the latest technology and the other is riding a donkey.”
The AI revolution will clearly benefit film-makers, but what is in it for us viewers? One potential drawback is that AI eliminates not only financial risk but creative risk, too. The fear is, if you fed in a vaguely challenging or experimental or atypical project into the machine – say Mulholland Drive or Under the Skin – the algorithm would discourage you from taking the gamble. Why not do a Dwayne Johnson action comedy instead?
Added to which, the data these algorithms are processing are actually human beings, who are inherently erratic, fallible and unpredictable. Your asset might go into rehab, get divorced or decide to go off and make shoes for a year. In 1996 Robert Downey Jr was arrested driving his Porsche naked through Los Angeles throwing imaginary rats out of the window. What algorithm would have recommended casting him as Iron Man?
“Already we’re seeing that we’re getting more and more remakes and sequels because that’s safe, rather than something that’s out of the box,” says Tabitha Goldstaub, a tech entrepreneur and commentator who specialises in artificial intelligence. Her work has raised deeper concerns about AI. Far from being a dispassionate tool, it often reflects the biases and prejudices of its creators, she says. “A lot of people think it’s maths so it can’t be biased, whereas in fact it’s completely the opposite way around: it’s maths, and therefore it’s data, and whatever data you feed a machine will have bias in it. The world is biased, and so these machines exacerbate our own biases.”
She points to the fact that no women have been nominated for best director at this year’s Oscars, and only five women have ever been nominated. The AI might well conclude that if you want to make an Oscar-winning movie, don’t hire a female director. “You can’t help but think: do you really want to leave these people in charge of creating an algorithm that predicts the next hits? It’s quite scary.”
But there are signs that AI could help reduce bias. The producer Todd Garner cited an example on his Producer’s Guide podcast recently. Garner’s company produced the Marlon Wayans comedies White Chicks and Little Man, and did very well out of them. But when it took a new Wayans project to traditional studios and financiers, nobody bit, he says. “Because there was no way, they thought, that this movie would play internationally … so it’s a purely domestic play [a US-only release]. Netflix came in and scooped it up immediately, gave us a lot more money to make the movie and the movie was huge for their site. Because Netflix had hosted White Chicks and other Wayans movies on their site, they knew that Marlon Wayans has a worldwide audience. If you don’t have that data, your instinct is, well, African American movies don’t play overseas, period.”
Now that AI is here, there is also the question of how much further it will infiltrate film-making. Screenwriting, for example. Things have moved on since Sunspring in 2016. Last year, Kevin Macdonald directed a one-minute advert for Lexus entirely written by an AI – that had been trained on 15 years’ worth of luxury ads.
As well as analytical tools, ScriptBook is developing a screenwriting AI, which it has given the dystopian sci-fi thriller name of Deepstory. “It really is a co-creator,” says Azermai, reassuringly. “We envision a next-generation writers’ room where whenever they don’t know where to head to for the next scene, they would have Deepstory create it. The engine takes into account everything that you’ve written, and it will deliver you the next scene, or the next 10 pages, or write it to the end.” Azermai admits Deepstory still has much to learn. “The consistency in writing stays for another 10 pages and then the AI becomes a bit crazy – sometimes it kills the protagonist for some reason – but it’s improving. Within five years we’ll have scripts written by AI that you would think are better than human writing.”
The movies themselves have trained us to be suspicious of AIs. They make terrible substitute children (Spielberg’s AI), and even worse romantic partners (Spike Jonze’s Her). They often decide humans are expendable and take over (Kubrick’s 2001) or they enslave us in some futuristic dystopia (Terminator, The Matrix). Could it be that by embracing AI, Hollywood is unwittingly building an equivalent of the Terminator movies’ Skynet? Or, in a slightly less apocalyptic scenario, are we looking at a future where movies are personally tailored to individual viewers? Might we be approaching a sort of Turing test for machine-written fiction? Maybe we have passed it already and we just don’t know it yet. Sounds like a plot for a great, metatextual Hollywood conspiracy thriller. If we got Dwayne Johnson it could be a smash. Let’s crunch some numbers!