Things Will Get Shittier: AI and the Dial of Business Basiness
By J. D. ConnorJuly 22, 2023
SECRET INVASION, the new series that debuted on Disney+ in June, was supposed to be the Marvel Cinematic Universe’s Andor, a serious, adulty, prestigey entry in the late-hegemonic metafranchise. It’s not. Almost everyone is underplaying to the point of boredom, and Samuel L. Jackson and Olivia Colman are responding by trying to have too much fun. But if Secret Invasion isn’t Andor, it is the—the—landmark incursion of artificial intelligence into the Hollywood creative system.
Secret Invasion tells the story (so far) of Nick Fury’s attempt to head off the takeover of earth by a terrorist cadre of dissident Skrulls, aliens who can convincingly shape-shift into human form at will. The opening gambit is that the invasion has already occurred. There are a million Skrulls hiding among the humans on the planet, most of them chill yet traumatized, a few scheming to relieve humanity of responsibility for the planet by, I guess, killing us all.
But when Secret Invasion dropped, the big news was not that we were getting to know a deeper, richer version of Fury but that the show’s title sequence was done by AI. Or, not “done by,” but, you know, kinda-sorta done by AI and overseen by artists at Method Studios, the firm Marvel contracted with to make the sequence. Unlike the AI-boosted titles for HBO’s Westworld, the “AI” here looks like a slightly outdated iteration of Midjourney, half-edited against a grandiloquent backing track so that the whole thing seems very two-or-three-cycles-ago. In the context of the ongoing writers’ strike, where AI has been the most surprisingly contentious issue, the folks behind Secret Invasion were powerless to resist becoming part of a discourse they did not know how to handle.
They tried. Initially, director/executive producer Ali Selim was out front touting the magic of this thing: “It felt explorative and inevitable, and exciting, and different,” he told Polygon. None of that describes the actual titles, but it does describe his experience with the process: “We would talk to them about ideas and themes and words, and then the computer would go off and do something. And then we could change it a little bit by using words, and it would change.” This is literally the venerable process of giving and taking notes (“using words”), only with a black box (“it would change”) where we would ordinarily expect a person (“they would change it”). Regardless, Selim underestimated how swiftly the negative reaction would come.
Faced with significant blowback, Method Studios defended themselves like Nathan Thurm: it wasn’t just AI, there were other tools; it wasn’t about saving money, it was about the theme and the “desired aesthetic” (Skrulls!); we employed the same number of artists; the feedback process was the same; this is just art direction; still, we must note, “the AI component provided optimal results.” Never let it be said that Method slagged its AI vendor in public.
Yet both their statement and Selim’s are poised on the knife edge of incoherent boosterism: “explorative [?] and inevitable,” and, according to Method, “collaborative,” “iterative,” “leveraged,” and “intriguing.” The results will be, assuredly, “optimal.”
Just before Secret Invasion premiered, the Entertainment Technology Center at USC, where I work, held a daylong Synthetic Media Summit to discuss the state of the art in AI for the industry. The ETC is a separately constituted nonprofit inside the USC School of Cinematic Arts, and it is funded by major studios and industry tech providers, much like the MIT Media Lab, but on a far more modest scale and with more specific ambitions. The pitch is, essentially, a version of the Academy’s old Research Council, now nested in a neutral institution that serves as a feeder school for the industry as a whole.
Yves Bergquist is the director of AI & Neuroscience in Media at the ETC and is, for now, sanguine about the AI threats. Large language models (LLMs) cannot write scripts and won’t be able to “in the foreseeable future.” Currently, the models amount to what he calls “hyperscale parroting,” but to get to that hyperscale, they use an enormous amount of labor and computing power. They are “very expensive to train and operate.” OpenAI (i.e., Microsoft), Stability AI, and Midjourney are all currently operating at huge losses, but when it comes time to seriously monetize them, according to Bergquist, “it’s not really entirely clear to me that the unit economics of writing a ChatGPT prompt or a Stability or a Midjourney prompt make sense for the [entertainment] companies themselves.”
Cautions aside, the summit showed off nifty new tools for rotoscopy and offered some stunningly credulous accounts of new frontiers in “artmaking.” But the audience was most engaged by the discussion of copyright. Because copyright, for the moment, feels AI-existential.
Currently, AI-generated creative work is not copyrightable in the US because the Copyright Office has ruled that it does not have a human author. Does that mean the titles to Secret Invasion aren’t copyrighted? Clearly Disney believes they are. Disney has spent a century finding ways of taking public domain material and snagging the copyright for themselves, from Snow White to Pinocchio to the Little Mermaid. In the title sequence, the music is copyrighted; the order and display of the names are copyrighted; the edits between the shots would be copyrighted as an “original arrangement” if nothing else. But do all those copyrighted elements sit atop a slurry of public domain green goo? Could you extract, say, the image of Nick Fury and do something with it—something commercial? Something that wouldn’t be covered by fair use but that would still be legal?
I dare you. However AI-ish in their construction, those images surely went through some minimal processing to become copyrightable, although Method has not told anyone what those processes were. Any notable use of them is bound to get you sued by Disney, and if history has taught us anything, that is a brutal life to live. The point here is not that Disney is correct that they own those ownable images—I imagine they do; the point is that by cultivating a penumbra of uncertainty around the AI-ish image, Disney has extended its control over those images whether they are ownable or not. The aura of property has chilled your free expression. You don’t know if the image is owned; you don’t know if your friends are Skrulls. You have been not-so-secretly invaded. This is the strategy.
The Secret Invasion titles are a first volley from the creative-content oligopoly. On another front entirely, the major generative AI companies are all being sued for their use of copyrighted material as part of their training data. If the AI hype machine is busy distracting you with apocalyptic scenarios of malign superintelligences and whatnot, these lawsuits make the plausible contrary case: The malign effects of AI will be here as soon as we accede to the use of these models even though they are based on massive theft—the wholesale ingestion of copyrighted material. Deciding that such theft is acceptable because the results are “optimal” is, as attorney and designer Matthew Butterick has argued, authoritarian, and laundering violations of law through the black box of “emergence” amounts to granting “plausible deniability for everything.”
The upshot is that studios are going to do whatever they can to hold their own copyright—and to find ways to make their uses of AI copyrightable—while wielding the hammer of the gods against anyone who so much as Bings “AI montage art.” An anonymous writer howled in The Hollywood Reporter that, in ratifying the new DGA contract, “the directors let the studios just have the right to train up future AI on their work. In the battle of Silicon Valley’s plagiarism machines versus human artists, that is everything.”
Imagine that DisneyFox generates a local instance trained on the entire corpus of Simpsons episodes—we can even imagine a world where it “trains harder” on better episodes (episodes are coded as better if the following criteria are met …)—and imagine it has ingested all the facts necessary to be familiar with the references in that world. Both of those seem plausible, if wildly expensive. (Ingesting facts is not what LLMs do, but it can seem so.) Does that local AI at DisneyFox uncopyright all the jokes and stories it generates out of the copyrighted materials? If so, what if a prompt engineer/executive producer—call this new job “PEEP showrunner”—sifts through the uncopyrighted jokes and plots, keeps some, discards others, and tweaks a couple, before sending them off to be animated? Is that enough to recopyright them?
Imagine further, somehow, that this Simpsons script is now copyrightable. Is the script the work of the prompt jockeys? The prior writers? How many of them? If they reuse a joke, do they owe the writing staff that came up with it? Does the filtration and recombination process also need to generate a kind of R-squared value to attach various compensation metrics to previous writers based on their contributions? What if DisneyFox launched The SAImpsons, a meta-gag much less funny than that eerie Muppet Babies parody Tiny Fuppets or Space Ghost Coast to Coast but still built on the archive, however diffused? How opaque is the black box of AI in a world where it can generate property?
The immediate questions for generative-AI-generated media are not about Turingian possibilities of emergent subjectivity and run-amok “agents,” but about opacity, surveillance, and power. And the promise of AI—the one that it might deliver on—is that some prompt jockey can get the thing to perform “optimally.” As scholar and tech organizer Meredith Whittaker explained on the podcast The Dig, the aim is to replace the system of unionized craftwork with a “gig economy platform” where those same workers “have to come back in under the title of an AI editor with no stability, no long-term future, no benefits.”
The production of a movie or TV show gives rise to a nearly infinite series of questions of tone. At every turn, one needs to know how intense, how florid, how knotty, how campy, how fun, how luxe, how shabby, how ominous, how cheerful it should be. Everyone who might somehow contribute to that tone runs this gauntlet of decision (and, I guess, explorativity?). When the process works, you get a new constellation of nuances, an affective world of sufficient range and power that it merits revisiting. This is the promise of long-form storytelling. Twin Peaks: The Return, What We Do in the Shadows, South Side—you have your own candidates. When it doesn’t work, as in Prime Video’s Citadel, it can feel like the digital Stanley Tucci asset has been dropped in from another series altogether.
I am talking here about the Great Dial, and in American life there are two fundamental images that backstop this modulation of control. The first is Ralph Waldo Emerson’s, and it appeared in the introduction to the magazine The Dial in 1840:
And so with diligent hands and good intent we set down our Dial on the earth. We wish it may resemble that instrument in its celebrated happiness, that of measuring no hours but those of sunshine. […] [L]et it be such a Dial, not as the dead face of a clock, hardly even such as the Gnomon in a garden, but rather such a Dial as is the Garden itself, in whose leaves and flowers and fruits the suddenly awakened sleeper is instantly apprised not what part of dead time, but what state of life and growth is now arrived and arriving.
This is a utopia of registration, an almost unlimited confidence that by doing your thing as precisely as possible, you will find yourself in tune with the world. There is no friction; the feedback is instant; our dial would belong to the garden, as we ourselves do.
The other dial is @dril’s: “turning a big dial taht says ‘Racism’ on it and constantly looking back at the audience for approval like a contestant on the price is right.” Our confidence is shot; feedback is blunt; attunement is reduced to being in sync with a wretched culture to which we have brought only the prospect of optimized degradation and violence. This March 2017 tweet from @dril captured the early days of the Trump administration, but, as with the best of his work, it applied to every “centrist” Democratic accommodation to the supposed demands of the masses.
Hollywood executives have their own version of the dial that falls between the two. At its most extreme, it is an unwavering confidence that the quality level of a project can be adjusted to suit its budget and audience. That confidence is not always unwarranted. Back in the 1970s, when Barry Diller was running Paramount, the studio decided to release a recut, PG version of Saturday Night Fever (1977) after the R-rated version had completed its run. Diller explained:
It is not the same film because it doesn’t have the impact of that raw reality. Is it as good a film? I doubt it, but, at the same time, I don’t think people who go to see the PG version will be cheated. They’ll see the same story, hear the same music, see John Travolta—although he’ll be more like the Travolta they’ve seen in “Grease.”
Diller is not quite saying that the PG version is not as good, but it doesn’t matter because you pigs will like it anyway. Just as important as his judgment is his status as a figure of total awareness and control: he knows what good is, and he knows what good enough is. The job is to provide good enough entertainment to make as much money as possible. The PG version added almost $10 million to the movie’s haul.
Yet executives’ confidence that they can relatively simply ratchet the quality of a project up and down is thoroughly underwritten by the maniacal focus on craftwork in Hollywood. Everyone in the industry knows that good takes talent, time, and money, and when you announce that a project will be rushed or underfunded, the adjustments are both reductions and constraints. These craft standards are not simply free-floating commitments to a set of criteria for excellence. They take hold in an industry of precarious professionals who, with each job, are pitching for their next gig. Striving to transcend a production’s limits is a crucial skill. Get a reputation for working on garbage, and, well …
But if significant portions of craftwork are ported to a generative AI—or if the allotted time and money are squeezed on the assumption that the work will be roughed out by AI and fixed by reclassified pros—then there is less resistance, less of a quality reserve, less solidification of the local sense of what it takes to do a good (enough) job for this producer under these constraints. And things will get worse.
And that is the point. For the vast majority of us, over the medium term, the advent of AI at scale will be about making things worse. “Degrading the quality of the working conditions” will mean a “decline [in] the quality of the content,” as Edward Ongweso Jr. put it (on the Meredith Whittaker Dig episode). The practical challenge is not whether AI can replace creative human labor, or even whether it can serve as an aid to creativity, but whether substituting a shittier version matters. “Let’s make things worse and see if anyone notices.” The AI dial doesn’t just say “Racism” on it—though, to be sure, the AI dial definitely does say “Racism” on it—it also says “Crap.” Ongweso wonders whether the system can survive such a quality erosion, but for the executives in charge, there is no risk of that. If they overcrank the Crap dial, they can always recalibrate. Even if they pull it off, though, everyone suffers in the interim.
AI is already making things worse. You have probably seen some garbage crop up in your newsfeed. I was reading a local news story about some houses in Pasadena that had been spared the wrecking ball when the extension of the 710 freeway was blocked. It was illustrated not with a photo of the surviving houses but an (uncredited, uncopyrightable) AI version of something like “house in Pasadena.” Thanks to @generalslug, I now know about a sludge factory called The Enlightened Mindset that is churning out literally millions of “AI” “written” “articles” “illustrated” like this.
The AI-juiced search engines are feeding on these results, spiraling into complete uselessness. This spring I got pulled into my first AI plagiarism case: the student relied on a chatbot to generate plausible-seeming links to nonexistent sources, sources that were only being offered to support completely banal claims. (Student plagiarism breaks my heart because the products are so boring, so much more boring than these messy individuals actually are; it is such a waste of everyone’s time, an index of a system broken in who-knows-how-many dimensions. AI-enabled plagiarism is, again, worse.) “When AI comes for your job,” Josh Dzieza explains, “you may not lose it, but it might become more alien, more isolating, more tedious.”
Cory Doctorow’s term to describe the degradation of platforms as they progress from serving their users to serving their advertisers to ultimately serving only themselves is “enshittification.” The advent of generative AI at scale will not be that. Things will get shittier, but not because ChatGPT starts serving up discounterfeit merch like Amazon. And not exactly because the entire system will be built on occluded exploitation and copyright violation, although it likely will be. Things will get shittier because that is how the dial will turn.
When Richard Plepler was asked to quadruple HBO’s revenues to generate enough content to launch HBO Max, he said no, for the relatively simple reason that there was no way to make four times as much HBO-quality content in the system. So, AT&T found someone who would turn the dial down a little even if it meant spending a lot more. They nearly pulled it off. They also lost billions of dollars in the process. To get out of that hole, they sold the thing to Discovery, and you know how that has gone.
If AI is already making things worse, already prototyping its malign effects in news, education, art, and culture, it is doing so along the sluiceways those fields had built to accommodate themselves to decades of financialized austerity and political rot. Secret Invasion captures that too.
Skrulls can duplicate a human on sight, but to know their minds, they must draw out their memories using a version of a machine we first saw in Captain Marvel (2019). (It’s like a digital Pensieve from Harry Potter.) In the Secret Invasion rebel base, there are a dozen or so high-value captives imprisoned in vertical basket stretchers, hooked up to biomorphic memory displays. A Skrull with some technical skill can scroll through these memories. On their screens, the memories are blurry, gloopy, distorted, photoreal versions of the title sequence, and similarly to the title sequence, they are produced by a contracted VFX firm. Like the invading Skrulls, what Whittaker, on The Dig, calls “the computationally mediated degradation of labor” is already here, visible in the nonunion, shot-contracted visual effects and animation industries, in all those service firms that sit just outside the studios, beyond the reach of the guild agreements.
Meanwhile, over in the Indiana Jones franchise, our final dial, the Dial of Destiny—the Antikythera mechanism, a truly astonishing astronomical clockwork—has dumped Indy and a bunch of Nazis at the Siege of Syracuse in 213 BCE. Once the Nazis are defeated, Harrison Ford explains to his goddaughter Phoebe Waller-Bridge that he wants to die in Syracuse because that’s where history is being made and nobody likes him back in 1969. It’s some real dad shit where self-pity hides inside fascination, some real “what if they made the whole thing out of the History Channel?” (L’Ancient Alien, c’est moi.) He is wrong, of course, because PWB needs him, as do Karen Allen and John Rhys-Davies (in MENA-face as Indy’s old friend Sallah). How did he get to this place?
Using the dial, Mads Mikkelsen hoped to go back to 1939, kill that stupid Hitler, and win the war for the Nazis. But as their plane goes through the temporal fissure, Indy is screaming about how they won’t actually end up in 1939 but somewhere else because of “continental drift.” They arrive smack in the middle of the Siege of Syracuse, where they meet up with Archimedes, who realizes that the Antikythera mechanism works. At this point, PWB explains that the mechanism is not a time machine that lets you choose where to go but a “forced deck” where people from the future would always be going back to Syracuse.
And she is right. Because wherever they turn the dial of basiness, we are always headed back to the fight.
J. D. Connor is associate professor of cinema and media studies at the University of Southern California. He is the author of Hollywood Math and Aftermath: The Economic Image and the Digital Recession (2018) and The Studios After the Studios (2015).
LARB Staff Recommendations
Paul J. D’Ambrosio reviews Elena Esposito’s “Artificial Communication: How Algorithms Produce Social Intelligence.”...
Jeffrey Binder uses the rise of ChatGPT to explore the backstory of our unease with artifice....
Did you know LARB is a reader-supported nonprofit?
LARB publishes daily without a paywall as part of our mission to make rigorous, incisive, and engaging writing on every aspect of literature, culture, and the arts freely accessible to the public. Please consider supporting our work and helping to keep LARB free.