There are too many waterfalls here

Sora slop feeds

What are slop videos for? Any interpretation of them should probably start with their primary purpose: to serve as advertisements for AI companies. It seems pretty clear (if the preposterous comments by Sam Altman noted in this Verge piece are any indication) that OpenAI deliberately launched its latest Sora 2 model without heeding any concerns about privacy, copyright, or the software’s general abuse potential so it would cause maximum controversy and garner as much attention as possible, far more than a rash of 10-second cartoons would generate on their own, regardless of how hyperrealistic they are.

OpenAI’s apparently calculated assault on the legal and public sphere, and the outcry it has caused, has yielded the usual hand-wringing about how powerful and truth-destroying “AI” will be — and not the giant tech companies that have carte blanche to ignore laws and fundamental human dignity — and how everyone’s basic grip of reality will be revoked as people edit each other’s faces into various fantasies and nightmare scenarios. This all sets the stage for a discussion of how the power of “AI” makes its supremacy inevitable and we simply have to accept the greedy plutocrats promoting the technology to be the caretakers of our future, and not the flagrant enemies of the common good that they blatantly are. They can steal and hoard the world’s knowledge and mete it out in confabulated blobs to the subjugated and ever more ignorant multitudes until the data centers finally succeed in making the planet inhospitable for us all.

So much for the supply side, but what explains the demand, such as it is, for generated video? What allows ordinary people to become complicit in that soulless and life-negating project? There are plenty of obvious ethical reasons not to consume slop feeds or contribute to them: the indefensible energy waste, the corporate appropriation of the general intellect, the threat to people’s livelihoods, the inescapable reinforcement of bias and cultural stereotypes, the “slow cancellation of the future,” the complementarity with the “aesthetics of fascism,” and so on.

But even if you could suspend those concerns (and you can’t, they are built into the technology), I struggle to understand why anyone would choose to engage with slop, would find enough enjoyment in it to bother with it rather than block it and move on. Maybe my problem is that I am not social enough — I don’t participate in enough chats and social media platforms to need to post and react to a never-ending stream of low-stakes content that holds groups together. Generated content becomes, like ordinary selfies, acceptable fodder to circulate in social networks — ad hoc novelties for people to post when they don’t having anything particularly novel in and of itself to say. Ellis Hamberger argued that “Sora feels more like next gen Bitmoji to me than the next TikTok,” and that seems about right; it offers another tactic for keeping chats going but offers little in the way of general interest.

No doubt, slopfluencers will emerge who aggregate large followings on various social platforms for sharing well-curated generated clips — the ones that have some captivating novelty or weirdness to them, for the time being — but they will just be servicing a content niche, which is all that “the AI revolution” seems to have amounted to. Max Read suggests here that

we might look back on Sora as the moment OpenAI settled in and allowed itself to be fully annexed by the social-platform sector — the AI boom ultimately less a regime change than the minor origin story for the latest entrant into small club of mega-platforms minting money from targeted advertising.

In other words, generative models don’t lead to “artificial general intelligence” but simply more content to keep people using their phones (and keep them under tech-company surveillance).

A Sora feed and Sora-moji make more sense to me than chatbot “companions”: That people would use generative models to help them sustain social connections rather than replace them corresponds with my hopeful belief that the only thing that is ultimately worth anything to anyone is other people’s time and attention. Slop might be easier or slightly less risky for some people to share when they want to be present online but don’t feel like they have anything interesting to say.

But this doesn’t explain why people would want to consume slop, or find it more interesting than other material they could share. Are Sora clips really in competition with TikTok clips for people’s idle consumption time? Are they as effective at making people distracted and distractable? Do generative clips offer anything (beyond fleeting novelty right now) that TikTok’s algorithm can’t provide a more human version of?

It seems self-evident that generative video makes the world of representation (if not the world in general) more boring. One could hope it would re-enchant those forms of visual experience that resist simulation, but instead there is a sense of theoretically infinite video unleashing an infinite boredom, depleting our capacity to see by inundating our eyes with synthetic sights. Generative models mean that no one has to create anything they don’t care about, but they also mean that all media has more of that sense of indifference attached to it. And if we consent to consume it, it is because we are willing to internalize aspects of that indifference, taking some solace in it.

Often the mainstream discussion of generated video begins with the concern about convincing fakes destroying our trust in documentary media, but if anything, it would seem to increase dependence on institutions with sourcing protocols. When anyone can produce realistic-looking fakes, being a trusted institution capable of persuasively certifying documents becomes more valuable, and the power of social “truth-making” becomes even more centralized and more pronounced.

At the same time, it seems like most clips aren’t trying to be persuasive on that level but instead hold attention by tricking people for an instant and then becoming “fun” or “watchable” (rather than disappointing and pointless) precisely because they are fake. They aren’t necessarily framed or understood as news. The criteria for them has less to do with showing real events but with their being legible — they have to be appropriately targeted jokes that one is in on; they have to be worth the attention spent on them. As Ryan Broderick suggests, “the online platforms that created our new world, run on likes and shares and comments and views, reshaped the marketplace of ideas into an attention economy,” and generated videos compete in that marketplace, where documentary veracity is mostly insignificant. Now the attention economy has “untethered popularity from tastemakers — cultural, political, financial — and turned it into something nakedly transactional,” Broderick argues. Regardless of their facticity, or what the “tastemakers” want, or what trusted institutions bother to verify, popular videos automatically become “true” in terms by virtue of their circulation (the only kind of fact that matters) and how they shape widely received narratives.

A flood-tide of heavily publicized generated video could have the effect of denaturalizing any tendency to treat video as automatically documentary. In this “defense of generative video accelerationism” Byrne Hobart claims that

for the purpose of truth-seeking, video is worse than text, not because it necessarily misinforms, but because it leads to overconfidence. The way you consume text is that your eyes scan a series of symbols, which you mentally convert into words and then into concepts. The way you experience video—a moving image with synchronized sound—is exactly how you experience real life. So video always feels more tangible. And since your idea of what’s normal is a function of what you experience, the specific thing video can do is make some things feel a lot more ubiquitous than they are.

Implicit here is the idea that realistic video has been an expedient way to inculcate us with media companies’ ideas of what we should believe about how the world works — it stocks us with false experiences that confirm bogus narratives about the distribution and effects of power. “At this point, every day produces enough video content for 24/7 programming promoting whatever ideology you happen to have,” Hobart writes.

That reminds me of Yves Citton’s argument in Mythocracy about how “the imaginary of power” works to “script” people’s experience of what is possible (although the “scripting” metaphor sort of muddles the comparison):

societies can only orient their development according to the possible futures that their participants have been able to imagine (visualise, envisage, invent, dream). The imaginary of power is not, therefore, a ‘theory’ that comes along after the fact to provide the analytical explanation of the images circulating around us. Rather, it is a set of schemas that we experience insofar as we use them. It is a set of imagos, of ‘expansive forms’ (patterns, Gestalt) that shape our expectations inasmuch as we are able to reconfigure them. They are spectacles that help us see ‘reality’ only by filtering what we see of it.

Algorithmic social media (now abetted by generated clips) can provide these “schema” and “imagos” that narrow the “possible futures” that viewers can imagine. By why do they keep watching? What makes it pleasurable to have reality pre-formatted for us in this way?

At first I was thinking that generated videos would mean that one comes to every video, and perhaps everything in the world, with a kind of weary skepticism, the same mental armor the we have equipped for advertising or for algorithmic feeds. But people also like those things too and don’t have to be forced into watching them. The skepticism mingles with a willing suspension of disbelief, and with the yearning that consuming videos could really be taken to add to one’s personal experience.

Generated videos, like other formulaic forms of entertainment, promise consumption without concentration, offering a kind of pre-patterned perception. Rather than having to expend the semiconscious effort to organize sense data into concepts and relations and synthesize the manifold (as Kant puts it), one can ingest content that is predigested and schematized, algorithmically presynthesized by statistical associations between word sets and image collections.

The viewer can thereby watch ideology (as it has been encoded over the ages in the dominant discourse that’s overrepresented in datasets) simply unfold in front of them as material that feels immediately “true” or looks just like what is expected, even when impossibilities are depicted. They can take in the received ideas without having to actively think them. That way ideology is reinforced as something that is just there to be seen, like sense perceptions — which must be gratifying for us on some level, a kind of relief that everything that we are supposed to accept is there right before our eyes. It is like peering into a collective dream while disavowing responsibility for it.

Claire Wilmot, in a piece for the London Review of Books, calls generated images “fascistic dream machines.” She explains that “part of the misunderstanding of the deepfake threat stems from the idea that it is a problem of bad information, rather than a problem of desire (or the material conditions that shape desire).” Generated images “offer clear, illustrative diagnoses” of “alleged problems,” which, if you are willing to surrender to them, must be a relief to see, particularly when you just want distraction or compensation for your own personal share of grievance.

As with other kinds of direct entertainment, the consumer’s impotence — their incapacity to imagine a better world or the reality of different ways of being — is turned into a kind pleasure and compensation for itself. When consuming images, it may be that people don’t typically want information (which reinforces impotence); they want ideology (which “feels true” and requires no power of mind). Part of the appeal then is indulging in unfettered cynicism. Wilmot quotes one bigot who seems to revel in it:

A Londoner spreading deepfakes of white women saying they don’t feel safe ‘because of migrants’ told me impatiently that everyone knows the videos aren’t real, but I was missing the point: ‘It’s about us showing everyone what’s really happening.’

Generated video allows consumers to inoculate themselves against events and representations that don’t conform to their schema by instantly offering alternatives that soothe them and match their expectations: They can enjoy their own ideological interpellation as a movie, or an endless feed.

But also key to generated images appeal us the ease with which they bring ideology before one’s senses. It supplements algorithmic feeds’ tendency to make the worldview tailored to one’s consumption patterns seem ubiquitous and self-evident, beyond question.

If you have to work to imagine the images to confirm an ideological distortion of how the world works, you begin to lose the libidinal benefits of that worldview, which grant a kind of identity that requires no work to articulate. Whereas the fact that a machine can generate these “true-feeling” images instantly confirms that the images can do what they are supposed to do, which is to protect us from thinking. Generative models are well suited to articulating ideas and wishes that people don’t want to have to make the effort to think through themselves, because the “pleasure” (such as it is) in them is in their apparent automaticity, in the speed with which prejudice formats and pre-digests and orders the world.

Video generators allow people to experience ideas or beliefs as content without their having to invest their imagination into making them real, into “really” believing in them and coming to terms with the implications of their beliefs. They are more like gaming engines than truth simulators.