Over recent years, we've seen artificial intelligence systems designed to write software, compose music, paint works of art, and even pen news articles, but the machines have been notably quiet in the medium of fiction storytelling. Designing an A.I. system that can write the screenplay for a movie, or compose a great novel, has posed a big challenge for researchers. So just how close are we to having machines pen our blockbuster films?
Sick of Ads?
Join more than 500 New Atlas Plus subscribers who read our newsletter and website without ads.
It's just US$19 a year.More Information
In June, a bizarre short film entitled Sunspring premiered. The film starred Thomas Middleditch (Silicon Valley) and chronicled the cryptic love triangle between three people inhabiting a strange futuristic office. Filled with incoherent non-sequiturs and inexplicably surreal tangents, the film could be considered either a compelling dream-like fugue or an amateurish mess.
In actuality, this odd film is something much more interesting. Sunspring is the first completely A.I. penned short film, developed in a collaboration between filmmaker Oscar Sharp and A.I. researcher Ross Goodwin. Appropriating a general text-recognition A.I. algorithm, Goodwin fed his system scores of science fiction screenplays from the 1980s and 1990s. Using films like Ghostbusters, Bladerunner and every episode of TV's The X-Files as its inspiration, the machine learned how to communicate in screenplay format and composed Sunspring.
The resulting film turned out to be stiflingly incoherent with literally no clear structure, but it revealed several fascinating recurring cinematic tropes that the A.I. seemingly recognized across numerous screenplays. Characters are constantly exclaiming confusion, with "I don't know what you're talking about," becoming a chorus-like refrain, while Thomas Middleditch's character shouts "it's not a dream" recalling countless reality-bending sci-fi stories.
The clear takeaway from Sunspring is that we're a long way from making a decent, or even coherent, A.I. penned movie. While A.I. systems have been developed to generate impressive musical pieces or works of visual art, it seems that fiction storytelling is a much more complicated beast. With its delicately complex combination of narrative, character, dialogue and structure, it seems like one of the critical barriers in developing a true form of artificial intelligence. A.I. pioneer Danny Hillis summed it up when he said, "The key thing that will make [artificial intelligence] work and make it acceptable to society is story telling."
A website called CuratedAI recently launched with the intention of acting as a repository for A.I. generated poetry and prose. Site founder Karmel Allison, a San Francisco-based software engineer, created a neural network algorithm designed to compose original machine-written pieces of poetry. Named Deep Gimble I, the algorithm has been loaded with a vocabulary of 190,000 words and its work is currently featured on the website. The poetry and prose the A.I. generates is definitely verbally discordant, but it does compellingly mimic the cadence and rhythm of classic work in its medium.
Earlier in 2016, a novel written almost entirely by an A.I. system passed the first selection round in a Japanese National Literary competition. Titled The Day A Computer Writes A Novel, this meta-tale had its human overseers direct the plot and characters while the A.I. generated the actual sentences. One judge described how the novel's ultimate shortcoming lay with its character descriptions, but the overall result suggested that a degree of human oversight or participation could make this kind of A.I. written fiction actually work.
The full feature
Impossible Things is the first attempt at an A.I. generated work of feature film storytelling and embraces the idea that having an human hand working with the machine is necessary. Mathematician Jack Zhang spent five years creating an A.I. that analyzed thousands of horror film plot summaries alongside their corresponding box office results. The idea is that the system can create a series of plot points that reflect popular audience tastes by crunching a serious amount of data. Understanding the limitations of the technology, a human writer was recruited to take the A.I. generated premise and plot, and add structure, dialogue and character.
The subsequent screenplay, Impossible Things, is now waiting on Kickstarter funding to move into production, but Zhang and his team have released a short summary of the A.I. co-written film and it predictably reads like an epic conglomeration of every horror movie cliche imaginable. The teaser trailer created to support the Kickstarter campaign, which was also written by the A.I. suggesting key things that would appeal to the film's ideal audience demographic, gives an appropriate indication of how unoriginal the results this kind of data-mined A.I. system can be.
Every single idea, shot, and audio trick can be traced back to prior successful films in the genre. Creepy kid singing a nursery rhyme? Check! Sound of a door creaking? Check! Ghost-like figure walking with a bloody knife? Check! It plays close to a parody of horror film tropes and in no way displays the whip-smart reflexivity that filmmakers like Quentin Tarantino deploy when they experiment with similar classic genre tactics.
The unsettling idea raised by this form of A.I. generated filmmaking is that the technology can only ultimately feed back ideas we have enjoyed in the past rather than creating fresh, novel and meaningful new juxtapositions. Judging by the content of most modern big-budget Hollywood cinema, this method of generating content, by replicating past success, is already a path many movies already tread – Star Wars: The Force Awakens, I'm looking at you!
Netflix is the power-player with this form of data-driven content generation, and while it doesn't have A.I. systems literally creating its product, it does tailor all of its in-house productions to identified audience habits. Netflix has a treasure trove of data at its fingertips and is able to understand the viewing habits of its audience in ways no media producer ever has before. Netflix not only knows how quickly you binge through a series, but it can track the minute you stop watching an episode, if or when you come back to that show, and how you navigate through content in its library. This data allows it to mould its own productions to their audience's preferences.
Back in 2012, before the debut season of its first original series aired, executives were transparent about how they were using mined data to conceive original content. Steve Swasey, VP of Corporate Communications said to GigaOm at the time, "We can look at consumer data and see what the appeal is for the director, for the stars and for similar dramas".
This strategy has obviously been successful. Regardless of the creative or critical success in Netflix's productions, the audience demographics of each project has been incredibly clear. Its most recent series, Stranger Things, is a prime example. One can so visibly see how an algorithm could suggest that a show with those parameters would be successful. Do you like 80s movies and have a nostalgic connection to them? Are you a fan of Winona Ryder or Stephen King novels? The series, while creatively only mildly successful, is a supremely well-executed mash-up of current hipster nostalgic obsessions from Steven Spielberg's Amblin films (E.T, The Goonies, Close Encounters of the Third Kind), to its synth-heavy score reminiscent of John Carpenter horror films (which are referenced frequently) and its retro typeface credits that could be ripped from the cover of a Stephen King novel.
Not in 4,500 years?
So where does that leave us in the world of A.I. generated storytelling, particularly in the realm of film and television?
Film and television are still such complex creative mediums with long gestation periods from pre to post production and a large assemblage of persons involved across that production process. As we saw with a fully computer-generated screenplay in Sunspring, A.I. currently has little to no understanding of the nuances in character development and lacks the ability to build a coherent, meaningful narrative structure. At the other end of the spectrum, with productions such as Impossible Things, we simply seem to get A.I. assisted, data-mined compendiums of cliches blindly mashing up ideas that worked in prior financially successful films. Netflix also surfs that line of data driven content production, and while it has had its own volume of creative hits and misses, we still can't shake the feeling that this mode of production is frustratingly uncreative.
In eccentric filmmaker Werner Herzog's latest documentary Lo and Behold: Reveries of the Connected World, Stanford A.I. researcher Sebastian Thrun mentions to Herzog the inevitability of a machine at some point being able to make a film as good, if not better, than Herzog. The 73 year-old, self-professed technology luddite, bristles at the statement, replying, "Absolutely not!" In interviews Herzog has reiterated stoically that not in 4,500 years could a machine make a film better than he. The statement is steeped in classic Herzogian arrogance, but it does allude to something fundamentally human present in the process of creating meaningful fiction.
A.I. may be currently able to offer reasonably interesting simulacrums of original content in other artistic mediums from music to poetry, but it seems that fiction storytelling is a tougher mountain to climb for machine mimicry. It's one thing to use data mining as a way to generate a fictional narrative, but in longer, more immersive forms of media such as film and television, it becomes stiflingly apparent if a work is either mindlessly incoherent or discouragingly derivative.
A great film speaks to its audience and offers a perspective on the human condition in ways that are often abstract or allegorical. We can be given a unique insight into the world through someone else's experience and this results in a narrative generating its meaning and impact in ways that often cannot be quantified.
No matter how much data a machine can crunch, will it ever be able to offer us a meaningful and effecting perspective on humanity? Can it generate insights into our experiences in this world that are new or profound and then communicate those ideas through a fictional narrative?
These are the hurdles A.I. developers currently face, and while it probably won't take 4,500 years to overcome these barriers, they certainly pose some fundamental questions about when and how a machine could develop independent and creative thought. In the meantime, we can just console ourselves with the latest Hollywood reboot, remake or sequel, and realize that machines are already pretty well represented in Hollywood.View gallery - 2 images