The Last (AI) Picture Show
The first in a series of short pieces on artificial intelligence and visual culture
Beginning with a confession: I want to see the AI-generated film that would supposedly end the film industry. This would be the film that puts all of Hollywood out of work, a film in which an AI assumes every creative role, replacing the thousands of jobs that make the movies an engine of capitalist plenty. An AI would write the screenplay, generate the CGI actors, sets, costumes. It may even design the distribution strategy and marketing. After this film is released, there would be no more need for gaffers, no more need for art directors, set designers, anyone. With this near-magic technology, reality would no longer be an obstacle, and the dream machine would cease to be a metaphor.
Parts of this technology already exist. Actors are digitally resurrected in almost every blockbuster. Concept artists seem to be already using text-to-image generators for their design ideas. Studios use software to predict whether a screenplay will be a hit. But despite these glimpses of an automated AI future, we still need humans to write films, direct actors, and build sets. The idea of an AI capable of making anything worth watching is still notional, if not impossible. As of this writing, AI cinema is the stuff of Instagram accounts and post-internet artists. But that hasn’t stopped two unions from shutting down Hollywood, in part, out of a fear that AI may one day devalue their members’ work. Their caution is appropriate, even if it is premature.
Six years ago, I published a short story about artificial intelligence replacing the Hollywood studio system. In the story, “The Making of The Entire City,” the film industry has collapsed, and executives turn to artificial intelligence to write a film that will rescue them from bankruptcy. After several years of research and testing, the AI writes a screenplay its makers claim will be a hit. The only problem is that the screenplay is thousands of pages long and is narratively incoherent. Ignoring their reservations, the studio execs hire a journeyman director and make the film, slavishly following the nonsensical screenplay. The resulting film has a five-hour running time and owes more to Stan Brakhage than Michael Bay. The film is also, against all expectations, an enormous box office success. Hollywood is reborn.
When I published the story in 2017, it seemed unlikely that Hollywood would turn to artificial intelligence to solve its creative and financial problems. Six years later, the premise is almost unimaginative. I wasn’t interested in predicting the future, nor did I think AI would advance so rapidly. Instead, the short story was partially a response to the work of artist and programmer Ross Goodwin. In 2016, along with director Oscar Sharp, he made Sunspring, a short film whose absurdist script was authored by an AI trained on several decades of sci-fi screenplays. During the film’s off-kilter seven minutes, characters regurgitate plastic eyeballs and say things like, “Well, I have to go to the skull.” Despite its absurdity, or maybe because of it, Sunspring became a viral hit, receiving millions of YouTube views.The filmmakers quickly made another AI-authored film, It’s No Game (2017), starring David Hasselhoff. That film’s plot? A Hollywood writers’ strike, naturally.
There is a moment during the development of a generative AI when the imperfect algorithm produces consistently surprising, oftentimes erroneous results. This is usually when the software is first released to a limited public—a brief period of useful imperfection, at least for an artist willing to exploit it. As anyone who has worked with generative AIs knows, “wrong” AI results are oftentimes more interesting than the correct answers, as they can contain a strangely nonhuman creativity. One example of this AI creativity could be found in GPT-2, the not-very-good predecessor of the widely used GPT-3. Most of what GPT-2 produced was more coherent than Sunspring’s screenplay, though it did make many linguistic “mistakes” by creating neologisms, broken grammars, and Dadaist solecisms. Early versions of Stable Diffusion produced similarly strange and artistically useful results—most famously, mangled hands and bodies. The subsequent “improved” versions of these AIs were far less interesting, creating a kind of instant techno-nostalgia for a software that was barely one year old.
Sunspring was an experiment. It was not meant to be a “great” film, by any measure. Its popularity was surely due to its uncanny novelty. But whatever its qualities, Goodwin’s AI did not generate a Hollywood classic. A “better” AI would probably converge on a screenplay that was more realistic. My story came out of speculating about that opposite could also come to pass: a popular AI film that looked unlike no other. Could a machine discover in the latent space of all screenplays a new kind of film that humans had overlooked?
The idea is not farfetched. In aerospace engineering, AIs are designing new kinds of hardware whose forms look truly alien. Then there is the work of AlphaGo, the Go-playing AI that beat world champion Lee Sedol.1 During and after its matches with Sedol, AlphaGo developed new styles of playing Go—a game that enjoys 4,000 years of complex strategy. Among other innovations, AlphaGo began to use unusual opening moves that would not normally be attempted by a human player. This was a rare example of an AI overcoming human bias—exactly the opposite of what many researchers are finding to be the case in other AI algorithms. I thought that, like AlphaGo’s innovations, perhaps an AI could discover a film somewhere in the nearly infinite field of possible films that is both crowd pleasing and extremely formally innovative.
Since I published the story, generative AI has unfortunately moved in the opposite direction, converging on the already known and the distressingly biased. The texts written by the latest GPT version are indistinguishable from the homework of a mediocre college student. This might be a great achievement for engineering, but it is hardly on the level of AlphaGo’s discoveries. Generative AIs such as Midjourney can execute perfect forgeries of science-fiction illustrators, though they have yet to produce much of original value. At best, much of generative AI remains a questionable augmentation of all-too-human artistic practices. See some recent headlines: “How I’m using AI to write my next novel” (Vox); or “How independent writers are turning to AI” (Verge); or “How AI will augment film production” (Variety).
Meanwhile, I will have to wait to see my fictive, perhaps impossible AI film. At least in the near future, there will be no AI Brando starring in the latest AI Huston film written by an AI Robert Towne. AI is incapable of worldbuilding, of creating a total, imagined environment that audiences would want to immerse themselves in. Admittedly, my fictive film is a willfully naïve reading of the future of AI art, whatever it may be. AI will “progress,” if that is the right word, more in fits and starts, sometimes subtly and other times all at once. As with AI-guided GPS and driving, AI art will slowly seep into the culture: first assuming one task, and then another, gaining control gradually and then suddenly. What is gained and lost, remains to be seen.
I wrote a piece about Sedol, randomness, and creativity for the 53rd and 55th issues of Mousse Magazine. It goes into these ideas in more detail.