Do Androids Dream of AGI?

1.

The dream of automating work and creating artificial life is at least as old as Aristotle and the myth of Galatea, but only after the Industrial Revolution did it become a prominent concern in the public imagination. For over a century now—if you start from Karel Capek’s RUR—countless popular fiction narratives have performed or ramped up the anxieties and dangers of producing artificial life or sentient machines. Or if you start from Shelley’s 1818 Frankenstein, with its organic-inorganic monstrosity, it’s been two centuries. 

In the more popular and influential works from the genre, these anxieties tend to be expressed through a familiar ideological vocabulary more concerned with humanistic pieties and nightmares of insurgent machines than the concrete political economies of digital technology and automation. From Blade Runner to 2001, from The Matrix and The Terminator series to Her and Ex Machina, much of the most influential fiction about intelligent machines consists of tonal variations on the fear of machine rebellion or the Christian tension around creating artificial life. These themes reek of misplaced class anxiety and stale misogynistic fantasies. In these narratives, the drama often centers on the threat machines pose once they become self-conscious, or on the difficulty of accepting their emergent humanity—and reckoning with how much less special this makes us, of course. Not surprisingly, many of these same dramatic structures are playing out in the most common reactions to the generative AI boom.

Systems of information retrieval and synthetic content generation certainly carry many social dangers, but none of them stem from the possibility that inhuman artificiality will defile and corrupt the sanctity of a perfectly organic humanity. To paraphrase feminist historian of science Donna Haraway, humans are and have always been cyborgs, in the sense that the species has emerged historically in a permanent state of open symbiosis with natural and artificial environments through tools and instruments of its own making. There never was a pretechnical human to be corrupted in the first place. Furthermore, we could say that large language models (LLMs) are profoundly “human” in their constitution since they are comprised of layer upon layer of alienated labor, from the underpaid workers in the Global South who are still integral in correcting and guiding the training of these algorithms to the hordes of writers and artists whose work is illegally pirated by these corporations to produce proprietary technologies for private profit. Made from different scales of exploitation across domains of time and space—from the shamelessness of their data scraping to their endless hunger for electricity—LLMs seem to express and perhaps ensure the consolidation of an all-powerful tech oligopoly, with its expansive empire of intellectual property and its (not-all-that-speculative) dreams of total surveillance and control. This is why the growing and uncritical mass adoption of these tools is a looming threat, and not because some sacred line is being crossed in God’s ledger, or because humans are somehow less special now that a neural network can produce a version of Pulp Fiction made by Wes Anderson. 

Much popular fiction has trained us to either accept the coming “humanity” of the robot or worry about the robot-slave taking our place, in an ironic twist of Hegelian self-consciousness. But these narratives have not adequately taught us to fear or even consider dystopian scenarios in which automated systems that have not reached anything close to perfection or self-consciousness are given full control of an already technocratic society, even though the latter scenario seems more likely to come to pass—indeed, we may already be living in a preliminary version of such a world.   

The French philosopher of technology Gilbert Simondon proposed that we fear the robot because we only know how to see and treat machines as if they were mechanized slaves. He also suggested that as long as our technological imaginary remains deeply tied to this vertical perspective of unilateral command over involuntary labor, we will only ever know how to build systems that amplify and reproduce new kinds of servitude. We all know what we’re supposed to do when the robots finally wake up to their miserable existence; we’ve all seen the movies. But what happens if human-made algorithms take over without ever waking up, creating a planetary-scale hell on autopilot—all too human, all too inhuman?  

Ex Machina, directed by Alex Garland, 2011, film still.

2. 

The first artificial neural networks were proposed by Warren McCulloch and Walter Pitts in 1943. Contemporary versions are essentially grandchildren of an algorithm for supervised learning first presented by Frank Rosenblatt in 1957, called the “Perceptron.” The kind of “connectionist” architecture favored by Rosenblatt was, in the decades that followed, largely abandoned in favor of Marvin Minsky’s “symbolic” strategies. Whereas connectionist architectures use a bottom-up approach to try to emulate the functioning of the human brain (hence the phrase “artificial neural network”), symbolic architectures use a top-down approach and a rule-based logical system to try to emulate human semantic reasoning. 

For many, the early dominance of symbolic approaches was responsible for the decades-long “AI winter” that followed the initial emergence of artificial intelligence technology. During this period, there were nonetheless advances in techniques like backpropagation and adversarial networks, as well as hardware innovations that facilitated parallel processing. It was only in the 2010s, however, that deep learning neural networks started to kick into high gear with the growing use of big data, which allows for “machine learning” based on statistical inference. In 2012, Google Brain first trained a neural network to recognize cats by feeding it massive numbers of cat videos. In 2017, a new “Transformer” architecture critical to deep learning was presented by a team of eight Google researchers in the famous “Attention Is All You Need” paper, which has been listed by Nature as one of the most-cited studies of the twenty-first century.1

Neural networks essentially use unfathomable masses of information to “train” computers to recognize and reproduce patterns across different domains. Their rise is inextricably linked, both causally and ideologically, to the consolidation of major tech platforms, especially Google. It was hard not to be impressed by ChatGPT when it was released, as well as by image generators like Sora. For a while it really did seem like they would rapidly get better and better. But even before the underwhelming release of the latest version of ChatGPT (and the termination of Sora), there were signs that OpenAI was running out of useful training data to scrape, and that this whole strategy is facing diminishing returns after the early waves of massive improvement. What’s more, the steep rise of synthetic content on the internet seems to be corrupting the training data for future models. 

Most specialists who do not have a financial stake in the industry have been skeptical of the extravagant claims about the future of AI made by tech CEOs and other Silicon Valley luminaries. But even if AI models fail to significantly improve in the near future, we can rest assured that this will not stop waves of opportunists from foisting AI “solutions” on nearly every sector of the economy. We can also imagine that many decision-makers in the public and private sectors will gladly adopt any shiny new way to avoid paying decent wages. Considering how aggressively these solutions have already been promoted and adopted all over the world in the last few years, it would seem that the AI bubble is here to stay, even if it does finally pop or deflate, to paraphrase the media theorist McKenzie Wark. 

3. 

It’s been more than two years since this new generation of AI models has been allowed to run wild over a largely clueless population, inaugurating a gold rush in Silicon Valley that makes the previous metaverse and Web3 crazes seem perfectly lucid and proportional by comparison. Piles of cash and energy, almost unimaginable in their scale (and dramatically unconscionable ecologically), are being burned by many of the richest companies in the world to secure their dominance over artificial general intelligence, and consequently over the rest of humanity, or so they believe. The stakes could not possibly be higher. 

What have been the social impacts of the technology so far? The whole structure of academia is being radically undermined by both the widespread use of chatbots by students, and perhaps more seriously, the eagerness to shove half-assed AI solutions down the throats of flesh-and-blood professors and educators. There are serious signs of rising unemployment in some sectors—like software engineering, illustration, voice acting, and translating—but it is still too early to predict how this will unfold over the long term. Hundreds of lawyers all over the world have been caught submitting legal documents that contain made-up references and precedents. Magazines, journals, and publishing houses of all stripes are being flooded with AI-generated pieces that are either nonsensical or painfully generic (some have published them alongside human content). The largest social media platforms—which were vibrant cesspools of automated activity well before OpenAI became a household name—are saturated with what has come to be called “AI slop.” The situation is particularly dire on X and YouTube. Google has started including an “AI overview” at the top of search results, guaranteeing a massive loss of traffic for newspapers (especially local and independent outlets) and other information providers, which had already been hemorrhaging ad revenue for decades thanks to Google and other platforms. This will almost certainly have destructive effects on an already fragile media ecosystem.

AI fruit slop.

It’s not an exaggeration to say that the most glaring social effect of the mass availability of generative AI as a consumer product has been an increase in grift, scams, and general inauthenticity online, as well as an overall decrease in the cognitive and creative quality of content. Most people over sixty seem to have no idea what is and is not real anymore. This is also starting to happen to younger generations more quickly than I would like to admit. More importantly perhaps, the AI gold rush has transformed the public posture of most major US technology companies from a mild concern with greenhouse gas emissions to a “drill baby drill!” approach to data centers and the energy they consume. And let’s not forget the growing number of people that have gone psychotic, even murderous and suicidal, by turning to bots as therapists, lovers, and confidants. 

To be fair, these AI models can produce convincing text and video, and they can simulate cognition in numerous scenarios. They can also organize and summarize vast quantities of data with impressive precision (although they are far from faultless). Despite their limits, AI models are already being used by many different professionals to assist in reasonable ways—although a 2025 MIT study suggests that for 95 percent of companies that have adopted AI, the productivity gains are negligible.2 It is crucial to note that these tools “hallucinate,” that is, they produce spurious results that they present as factual. The term is deceptive since it suggests that the machine is having a “real” perception when it produces adequate results and a “distorted” perception when it produces inadequate results; in fact, both generative processes are essentially the same for the machine, in a statistical sense. 

In short, these content generators can be very impressive—even eerie—emulators of perception and cognition, and truly awesome organizers of information. They cannot, however, reason in any meaningful sense of the term. They are not even remotely close to being sentient. Nor do they seem to be progressing in a linear fashion towards a general-purpose human-level intelligence. (Even the most delirious AI enthusiasts are starting to admit this after the dud that was GPT-5.) And yet we still see legacy outlets like The New Yorker and The Guardian take the possibility of AI sentience seriously.3 Prominent institutions still get millions to blabber about the dangers of a runaway AI destroying humanity, as if “alignment” were the greatest existential risk facing humankind. 

How does a ruling class and its attendant sycophants become so deeply alienated from the social reality that underpins their rule and the technological infrastructure on which their economic power is based? Can all these people really be this rich and this dumb? The answer is a resounding yes. But their dumb dreams do not come out of nowhere. Although these AI moguls and VC investors do basically live in an alternate reality—in bubbles shaped by fringe ideologies like “effective altruism” and “longtermism”—their silliest notions about AI come from widely shared social fantasies. The fact that digital technology constitutes the practical ruling hegemony of our society does not hinder it from also being one of the more profoundly obscure regions of society’s unconscious. 

4. 

What explains the massive piles of cash being burned by these companies is, of course, the dream of artificial general intelligence (AGI). Achieving this civilizational milestone would be, presumably, the only way they could ever turn a profit after the unfathomable waste of the last couple of years, and the further waste that is projected for the near future. The CEO of Nvidia recently estimated that by the end of the decade, overall investment in AI could reach $3 to $4 trillion.4 We should point out, however, that the term “AGI”—although widely used in the technical literature—is not clearly defined as a concept or a scientific goal. For some, AGI would be achieved when a machine developed human-like intelligence in virtually all cognitive areas; for others, it is an entirely new kind of recursive general intelligence that has the potentially infinite capacity to adapt and expand on its own. The most exaggerated descriptions evoke the divine. The fundamental vagueness of “AGI” doesn’t stop CEOs, engineers, and investors from throwing around the term like some kind of spiritual or ontological inevitability.  

As many people have pointed out—from the journalist Émile Torres to the philosopher Fabian Ludueña—techno-utopian dreams are often just secular transfigurations of familiar religious tropes. Many transhumanists are just disappointed Christians looking for a new daddy, or gnostics with mommy issues. Some of the most eccentric types say it quite explicitly. Asked if God exists, the most notable proponent of the singularity, Ray Kurzweil, famously answered: “Not yet.” 

Responding to these pathetic psychodramas avant la lettre, Simondon offers a different approach, arguing that automation should not be the ultimate goal of technicity. There are obviously countless tasks that can and should be automated; indeed, automation could conceivably redefine the relation between labor and technology entirely, helping to overcome mass exploitation. But the existing dream of automating everything away is ultimately about erasing all friction from social reality, achieving a state of total consumer passivity—even, ultimately, transcending the corruption of the body, like in Zuckerberg’s metaverse of floating heads and Musk’s Neuralink heaven. All the water in the world should be expended so that you don’t have to think a single thought by yourself ever again, and so that ChatGPT’s, Anthropic’s, or Alphabet’s golem can become your butler and your sex slave, if you can afford it. This is the end point of modern Western technology conceived as the total instrumentalization of nature and the birth of a new synthetic godhead, designed in sunny California.

5.

What new forms will social domination take in the epoch of so-called “artificial intelligence”? It seems too early to say, but the truth is that even before these models emerged, we already lived in a society with layer upon layer of automatism and inauthenticity. From high-frequency trading to algorithms that assist with judicial rulings and CV processing, from robocalls and drones to the ever-evolving arms race around spam and SEO, many social tasks already run on faulty automated systems. At the very least we can predict that these generators of derivative content and systems of automated decision-making will make a careless world even more careless. A brutal world of capitalist “mute compulsion,” corporate Muzak, and technocratic irresponsibility will become even more compulsive and brutal.5 

Integrating AI into the already entrenched structures of algorithmic governance and chronopolitical control will lead to a dramatic reduction of “the field of the possible to the field of the probable,” as the Brazilian media theorist Paula Cardoso has argued.6 The mass adoption of these models within economic and cultural production seems likely to turn the open-ended realm of virtuality and possibility that is human creativity into a dreary space of recursive statistical redundancy. Every built-in bias amplified, every structural prejudice reproduced, every blind spot multiplied.

Even I, a hater of these products, felt a hint of awe, even glee, when I first interacted with them. These systems really do signal the crossing of a strange border. Who knows what other weird artificial worlds might rise out of human ingenuity, if we ever get the time to really play around with them? While it is very difficult to predict anything with certainty in this field, we should at least consider the possibility that chasing the myth of the all-powerful anthropomorphic machine is in fact a criminally wasteful delusion that attracts extremely powerful individuals simply because they all watched the same movies and adhere to the same techno-triumphalist ethos. If the bubble bursts—or perhaps especially if it accelerates into something else—it will be nothing short of a disaster that such a wild venture into unknown cognitive territory was led by some of the worst people on earth, in such a sinister political moment. The dangerous entanglement between protofascist governments and reckless corporations guarantees that the worst possible choices are being made at every turn. The destructive synergies we saw at DOGE, or between ICE and Palantir, might be just the beginning. 

For Matteo Pasquinelli, the primary function of the current generation of AI is to abstract labor and social relations, transforming the collective accumulated work of the general intellect into proprietary structures of governance and control.7 Like any complex technology, but at an even deeper level, neural networks are the embodiment of massive collective efforts distributed over time and space, a crystallization of countless meaningful gestures that ultimately encompass the whole of humanity. It is singularly perverse that these architectures are developed inside organizations that are exclusively dedicated to increasing shareholder value at the expense of literally everything else that exists.  

The dominant imaginary of technology in general, and of digital technology in particular, is so profoundly alienated that even the oligarchic CEOs that rule us seem unaware of the infrastructural and ecological limits of their own systems of value production. It bears repeating: The fact that technology constitutes the ruling hegemony of our society does not hinder it from also being one of its most profound blind spots. No democratic or socialist future will ever be possible unless digital technology is liberated for the common good. We will never build a better world if we barely know how to dream of one.