Epic Forgetting

Below is a paper I wrote for (remote) presentation at this conference on the “Aesthetics of Democratic Life-forms” earlier this week. It’s probably not as dialectical as it ought to be, and it reprises some material from earlier newsletters. But having worked on it all week, I didn’t have the time or spirit left to write anything new for this newsletter, so I figured I would just post this instead.

1. The general anti-democratic nature of “AI”

Artificial intelligence, as its critics like to say, is neither artificial nor intelligent. It requires material resources and depends entirely on human labor, so it’s not artificial; and it has no intentionality or reasoning capabilities, so it’s not intelligent. It does nothing but identify patterns, classify data according to training-set labels, and optimize to minimize loss functions. Like so-called “smart” devices, it is technology put in position to automatically and algorithmically intervene in delimited environments.

“AI” is a euphemism that is meant to deflect attention from what machine learning actually is: not “the sudden spark of consciousness in silicon, but … a set of computational methods with political implications,” computer scientist Dan McQuillan writes in his book Resisting AI. AI is currently the most salient expression of what the late David Golumbia called “computationalism” — the idea that minds and computers are alike and that all thought is information processing, instrumental reason. “Despite being framed in terms of distributed power and democratic participation,” Golumbia argues, computationalism instead serves “the ends of entrenched power,” offering illusory and ideological experiences of individual mastery that mask the technology’s authoritarian implications.

In general terms, AI is the aggregation of massive amounts of data and processing power in the hands a few giant private companies to facilitate profit and coercive administration. “Whatever we think of specific AI applications,” McQuillan writes, “accepting AI means we are implicitly signing up for an environment of pervasive data surveillance and centralized control.” In specific terms, that pursuit of control means using data to develop predictive models, consuming vast amounts of energy and water in the process.

The data sets involved are necessarily large, and the makers of AI models openly aspire to totality — “to map out the entire world of objects” as Fei Fei Li, a lead developer of the ImageNet training set, said; or “to represent the collective imagery of humanity compressed into files a few gigabytes big,” as Emad Mostaque, the CEO of the AI company Stability, put it. Such a totalization presumes to transcend any subjective perspectives to offer instead what a uniform, generic “humanity” knows as reality.

Not only is it assumed that more data makes for better models and more accurate predictions, but since the largest machine-learning models are now being developed as open-ended in their capabilities — as “artificial general intelligence” — any kind of data is perceived as potentially relevant. It can all factor in to the billions of parameters that comprise their neural nets. It is assumed that with enough data, the formula for the entire objective world — the rules for how everything interconnects — could be reverse-engineered.

And accordingly, just as any form of data can be useful for a model, everything in the world and all the living experience in it can correspondingly be understood as nothing but data, capable of being abstracted from the conditions under which it is collected and made commensurate and comparable, immediately translatable into the same terms, as if all experience were fundamentally the same sort of content. AI takes an aggregate of representations of the world to be the world itself. It seems to demonstrate how, in Golumbia’s words, “the rationalist vision could be mutated into something like a full articulation of human society.”

As training data is amassed — typically through outright appropriation on an unfathomable scale from a variety of online sources — it must be categorized and labeled to work with machine-learning algorithms. This labor is often performed by low-paid workers who are treated as interchangeable, as generic humans. They are instructed to become as machine-like as possible in performing their annotations, implementing the decontextualization necessary for the models to use them.

The labeling process could be understood as a form of reification, and, as Adorno put it in a letter to Benjamin, “all reification is a forgetting; objects become purely thing-like the moment they are retained for us without the continued presence of their other aspects, when something of them has been forgotten.”

But when the world’s experience is converted to data, labeled, and assimilated in a predictive model, this forgetting itself is, in a sense, forgotten. This is because the model’s effectiveness at producing results, at making decisions, comes to supplant the forgotten material, the “other aspects” that situate experience socially, the embodied attentiveness to the world that constitutes our knowledge of it.

AI’s automated forms of decision making, regardless of their specific instantiation or situational impact, all contribute to a general understanding that non-datafied forms of experience are not relevant. They are ineffective: They are not part of the formula and have no hold on the world. Converting experience into data, abstracting away from any situated perspective, is what makes an experience real, consequential.

For AI models, meaning is predetermined as a statistical artifact and treated as inherently calculable. Everything is in principle already determined, and the possibility of generating new concepts — a collaborative, intentional human social process necessarily requiring the free play of our faculties — is replaced by more powerful machines identifying more and more statistical patterns, applying more labels, reifying more experience in advance.

The few companies large enough to administer these models become capable of implementing algorithmic decision-making systems across society as closed feedback loops, creating zones in which past data determine the future, reproducing historical biases, divisions, exclusions, and unequal distributions of power and opportunity.

As a form of “supercharged bureaucracy,” in McQuillan’s terms, AI black-boxes authority, conceals the bases for the decisions it implements, and impedes accountability for those decisions, creating what he calls “algorithmic states of exception.” AI models presuppose that decision-making should be a closed process based on statistical inferences, implemented from the top down, rather than a process of debate that draws on or builds shared forms of experience and elicits active participation. In this respect, AI represents the very opposite of democratic processes or democratic life-forms.

2. Generative AI as “democratizing”

Much of our experience with AI systems makes their centralized and antidemocratic nature plain: They are used to discriminate among people for benefits distribution, job searches and employee evaluation, credit and housing applications, medical triage, insurance claims, predictive policing, judicial sentencing, facial recognition, price discrimination, and on and on. You find AI anywhere a large institution can make individuals feel as though their existence has been reduced to their data.

Yet much of the current discussion about AI shows a different face, emphasizing how it puts tools in our hands to effortlessly expand our capabilities, placing us in control of machines that can endlessly produce at our behest. This discussion emphasizes the machines that can output text and images in response to natural language prompts — what is now usually referred to as “generative AI,” and which is coming to occlude the other, more explicitly coercive forms of computational automation.

Part of the ongoing onslaught of hype for chatbots (like Open AI’s ChatGPT, Google’s Bard, Facebook’s LLaMa, and so on) and text-to-image models (like Midjourney, Stable Diffusion, and Dall-E, among others) aims to reinvigorate the idea that the centralizing, surveillance-dependent technology of automated decision-making and imposed classifications can somehow be “democratizing.”

Thus text-to-image models are described as making artistic production accessible to anyone, without any need for the aptitude, time, or energy for developing and exercising craft skills. The company Stability, to take a typical example, described its launch of Stable Diffusion as “democratizing image generation,” “empowering billions of people to create stunning art within seconds” and, along with other generative models, bringing “the gift of creativity to all.” In OpenAI’s recent announcement of the latest iteration of its text-to-image model, it promised to “bring your ideas to life.” The implication is that ideas in your head are basically dead until technology animates them.

Technology can bring creativity to us as something reified, something we consume, and creative works themselves are presented as given, fully formed, ready to be enjoyed in a “culinary” fashion, as Adorno liked to put it — as an “accumulation of sensual stimuli” and not an “objective context of meaning.” (From Adorno’s Aesthetics lectures.)

In part, the reward for using a text-to-image generator is not the specific image it produces but access to that culinary mode of enjoyment, bypassing the demands of the sort of reflective judgment that aesthetic engagement requires. One can identify the concepts that went into the final product and treat that decoding as the essence of the pleasure it could possibly provide. We recognize patterns like the machine.

With generative models, making meaning is simply a matter of prompting. It doesn’t require careful attending to reality, or to the resistance that a particular medium of expression might present. Prompting is conceived as straightforward and accessible, yet at the same time, it is also presented as a new kind of “engineering” specialty, an open-ended form of coding in natural language in which one can develop expertise and secure a competitive advantage. It’s something one can master rationally, without having to make recourse to fickle imaginative abilities.

Large-language models are similarly seen as broadening access to information and skills and providing tutoring in every conceivable subject. They allow noncoders to code, nonwriters to write, nondesigners to design, and so on. LLMs offer the same reward as image generators, a way to endlessly and passively consume information as immediacy, regardless of its ultimate quality, which is of course notoriously unreliable. Thinking, in the context of LLMs, becomes a matter of making demands for information and consuming output.

Proponents of generative models tend to characterize this not as deskilling the process of thinking in general but as liberating the array of specialized skills from the control of the specialists, revealing all those skills to be no more than variants of information retrieval. Democratizing skills means turning them into the same skill. And this skill is less a way of interacting or touching the world than knowing how to manipulate data. In a sense, acting on the world is redundant, because it just retraces the pathways and connections the model has already traced. It would be like copying out a book that’s already been written.

Models supposedly level the field, discrediting elitist notions of creative or artistic ability. This belief can be seen in the bullying tone AI promoters often adopt on social media when they talk about how AI is “coming for” various jobs and people — taking their special abilities away. It shows as well in their general attitude toward art, which as critic Philip Maciak pointed out, treats it as “a problem to be solved … an uncomfortably irreducible remainder to optimizing the world.” AI proponents would like to reveal artistic creation to be a gimmick, and such notions as “genius” and “inspiration” — or the possibility of aesthetic judgment in general — as tricks, as “glib illusions.”

By making artistic production instantaneous, “genius” is supposedly exposed as a con, an ungrounded play for social status. AI models reveal that there is nothing distinctive about art, taste, or aesthetic experience, and no reason they should afford anyone any cultural capital. And once the cultural capital of these qualities is depreciated, they are revealed to be nothing more than useless, or at best just more information, another form of data to be instrumentalized by anyone with the means to do so.

People are thereby free to develop their own cultural materials on their own terms, unilaterally without interference from or deference to culture industry elites. By abolishing the mystifying presumptions and exclusions of aesthetics, generative models would seem to allow everyone to more freely enjoy the fruits of artistic production and consumption.

3. Aesthetics as necessary to the experience of freedom

But aesthetic experience need not be reduced to a squabble over cultural capital, and the idea of “genius” need not be understood in strictly elitist terms, but can be seen in a different register — in this case a philosophical register.

In the Critique of Judgment, Kant defines an “aesthetic idea” as basically the opposite of what generative models produce: It is a “presentation of the imagination which prompts much thought but to which no determinate thought whatsoever, i.e. concept, can be adequate, so that no language can express it completely and allow us to grasp it.” In other words, an aesthetic idea is untranslatable into a prompt, and a prompt will never yield an aesthetic idea. Aesthetic ideas, Kant writes, “arouse more thought than can be expressed in a concept determined by words.”

Aesthetic ideas are important for Kant because through them, he says, “we feel our freedom from the law of association” — we are able to have thoughts that are not predetermined in the strictly rationalized and instrumentalized way that an AI model determines its outputs.

Aesthetic ideas present “the imagination in its freedom from any instruction by rules.” An aesthetic idea “makes us add to a concept the thoughts of much that is ineffable … and connects language, which would otherwise be mere letters, with spirit,” which he defines as the “animating principle in the mind” — in other words, life itself, the quality of being alive.

Of course, it should be obvious that AI lacks this quality. Yet there is often an assumption that AI is edging toward being alive because of its apparent language competency. Or that it its somehow participating in the zeitgeist because it has absorbed so much human expression. But the essential point here is that geist is precisely what escapes datafication. Generative models can only ever simulate it; they can only ever substitute something fixed and reified for it. Rather than bring ideas to life, generative models kill them dead.

Aesthetic ideas supply the living component of language that links humanity together. Aesthetics points to the aspects of communication that can’t be autocompleted or directly instrumentalized, that are irreducible to concepts — embodied aspects including tone, gesture as well as intersubjective aspects like the context, idiom, iterability, and the particular common grounds between those attempting to communicate.

“Genius,” in Kant’s sense of the word, allows us to discover and manifest these aesthetic ideas that specifically derive not from rational cognition but from the “quickening of the cognitive powers.” It is not indicative of an elite skill that AI can redistribute; rather genius stands for a mode of resistance to the kind of rationalized dominance and entrapment in status quo meanings that AI systems impose.

As J.M. Bernstein puts it in a lecture on Critique of Judgment, genius can be seen as “the exemplary expression of human freedom.” And thinking about artworks, he argues — that is engaging aesthetic ideas — “is a difficult and exemplary way of thinking about the nature and meaning of human freedom” because “when we make artworks, we make original, new meaning.”

In other words, art is indeed, as Maciak put it, an “irreducible remainder” that impedes optimization. But impeding optimization is what’s necessary to leave open the space for new possibilities. Optimization is inevitably a formatting of the future to suit the established frameworks of the past.

Adorno would frequently discuss art in these terms — where art both a figures a possible way out of the fully administered society and the reification and exchangeability of everything but also the failure of that hope as artworks inevitably become commodified.

But for generative models there are no remainders. They can’t feel their way toward a meaning. They can’t have aesthetic ideas or make artworks in the sense that Bernstein means, so they can’t redistribute artistic capability. They don’t create works that figure freedom, or exemplify, in Bernstein’s words, “people making sense in ways that follow no existing rules of sense-making.” They don’t elude reification or hold out utopian promise that something can exceed exchange value.

Generative models instead negate all those conceptions. They depend on the premises that there are no “new and original meanings;” that all meaning is already determined and does in fact stem from existing rules, rules that are not intractable but can be systematically extracted from data; that everything can be exchanged into everything else at a higher level of abstraction; that no moment is irreducible or “nonidentical,” to use one of Adorno’s terms.

In Bernstein’s reading of Kant, “the notion of meaning cannot be reduced to the notion of determinate meaning”; he argues that scientific meaning — the ideal of fully determinate communicability, what AI models are designed to produce — is a “form of unwanted hegemony over communication itself.”

AI, and computationalism, in general, imposes that hegemony, insisting that all meanings can be communicated, and everything can be represented as data. But from the Kantian perspective, “the ideal of communication, of making oneself completely understood socially, requires the social acknowledgement of what cannot be made explicit in communication.” What is forgotten in reification must be retained or restored. “We have art,” Bernstein suggests, “because our ordinary communications fail to communicate what is most important to us.”

“In art,” he says, “the product — the meaning — is dependent on the process by which it is achieved. While in science — the other extreme,” epitomized by AI — “the process by which a meaning is gotten makes no difference whatsoever.”

When one engages with AI models, the meaning of the process is systematically precluded. If, as Bernstein puts it, “freedom is the capacity to act anew, and art is that human practice in which human beings routinely produce new, unique items,” then generative models are the opposite, a foreclosure of the possibility to create new meanings, new unique items — things that can’t be virtually anticipated by an algorithmic process.

Again, this seems like it would be obvious — you can’t get anything out of a computational process that you haven’t already put in. Math equations are not imaginative. The social process is what adds new meanings. Generative AI could be part of a social process — it can be used to make things that are then used socially — but the relation of a user to a machine in and of itself doesn’t make anything new. And the concern is the degree to which this relation could serve as a substitution for social relations.

It follows, then, that encouraging people to use generative models for feats of pseudo-creation can be a way of preventing them from experiencing freedom, from experiencing the specific kind of pleasure that Kant links to aesthetic judgment — the mind’s awareness of its own conceptualizing activity and how this capability authorizes a shared understanding of the world, a common sense.

AI models impede that awareness and offer an apparent compensatory gratification in the experience of convenience in and of itself, in a life that can be autocompleted for us, giving a sense of mastery without actual abilities to ground it. If aesthetics completes communication, AI prevents it. Or at best, it serves as a sort of dialectical clarification of how disappointing the scientific hegemony is, how suffocating is the experience of a fully rationalized and administered world, where everything is already conceptualized.

In consumer-focused forms of algorithmic prediction — the social media feed is an especially conspicuous example — output is oriented around an isolated individual user, and makes that individualization a key libidinal component. This is both gratifying and stultifying. AI-assisted composition and ideation — autocompletion — is similar: Our capacity for thinking is interrupted and steered away from the process of ideation, away from the engagement and harmonization of our faculties, and is directed instead toward trial-and-error engagements with machine-generated options, and a sort of consumerist pleasure in shopping among them.

We are limited to providing feedback on the machine’s decisions with regard to our own intentions, assessing how viable its abstraction of our supplied context is. The result is thought without any personal connection in it — purpose without purposiveness. Our relation to our own intention is deskilled, much as algorithmic feeds deskill our relation to our desires.

We don’t need to understand how to go after what we want or the social implications of it; we can consume our goals and desires as end products constructed and presented to us. Our consumption choices, our intentions, our life processes to the extent they are tracked and captured, are just forms of self-annotation, taming and ironing out our complexity to have ourselves rendered machine-readable, which becomes the key to thriving in an AI-assisted world.

Generative models are not democratizing by virtue of giving tools for expression to those who would otherwise lack the capability of turning their thoughts into reified objects. They don’t allow everyone to become artists, or make art more accessible or comprehensible to everyone, any more than cryptocurrency democratizes finance. They don’t constitute an assault on art-world expertise.

Rather they promulgate a consumerist position with respect to art, information, culture — a stimulus-response model that AI enacts and which the users of AI must adopt. They present the “social” as a data set that can be navigated without actual sociality. The models prevent us from experiencing the moments of imaginative spontaneity and free play that Kant located in aesthetic judgments, in conceptualizing without concepts, the imagination and the understanding harmonizing in a live, open-ended engagement with the world.

If language can be completely reified as math as AI presupposes, it could no longer express new possibilities. But it would remain useful for issuing commands. It could document what exists in pre-established terms, but it could not be used for dissent or for articulating negation, for imagining something other than what appears to be.

From this perspective, the models would “work” not to the degree that they synthesize plausible responses to queries (this wouldn’t matter because meaning wouldn’t really matter), but instead to the degree that they can be evoked and deployed to discourage people from using thinking freely in language, from sustaining language as an instrument of free thought. The ultimate aim of an LLM is not to make human beings more proficient in language but to make them be willing to do without language altogether.

The more that we can be convinced that AI is potentially useful and labor-saving, the more effective machine learning and surveillance will be at administrating an authoritarian society. And it may be that there is no middle ground — there are no “convenient” uses for generative models that don’t also augment their repressive possibilities. Generative AI would seem to be a way of restating AI’s “democratizing” potential, but in practice it extends the same reification and  bureaucratization into the aesthetic realm, foreclosing on the freedom that inheres there and supplanting the pleasure it provides with a pleasure in consumerism and the pseudo-mastery of instrumentalized control.