How to Survive Artificial Intelligence

The 1950s American pamphlet, How to Survive an Atomic Bomb, was more than a practical manual: it transformed the nuclear nightmare of mutually assured destruction into a sequence of manageable, individual actions, offering people the illusion of control and agency in the face of an overwhelming threat.¹

With artificial intelligence, the dynamic changes.² We are no longer facing a possible catastrophe, but are immersed in one already underway. In dominant, mainstream discourse, the narrative usually oscillates between three opposed and oversimplified poles: AI as the threat of machine domination; AI as a salvific promise, capable of remedying the chaos toward which capitalism is leading us; and finally, AI as a transhumanist tool for enhancement, destined to create an "augmented" elite.

Things are probably not quite so simple. It therefore seems necessary to initiate an inquiry capable of confronting the complex set of phenomena generated by this new context, in which AI has become part of our everyday landscape.

This essay seeks to lay the groundwork for a broader discussion, by adopting a specific approach and, for now, limiting the analysis to certain central lines of investigation. These lines do not exhaust the overall picture but allow us to begin sketching a path.

My approach is to address this technological leap politically, socially, and economically — adopting a perspective that takes quantum physics as the real framework of nature, thereby overcoming the illusion of a purely Newtonian model. In line with this quantum stance, I employ a "diffractive" methodology, weaving in references from historical materialism with those of new materialism.

The first step is to situate AI in its historical context. Like every technology, it does not emerge in the abstract but develops within specific historical, political, and socio-technical conditions, until it eventually infiltrates nearly every sphere of life. Yet, despite the massive investments and grandiose proclamations (from Trump downward), its capitalist economic model remains far from consolidated.

The main sections of the text will then develop an analysis of AI as a relational reality, articulated through two complementary perspectives. In the first, I examine the dynamics, methods, and responsibilities through which AI is modeled, constituted, and shaped: who builds it, in whose interests, within what power relations, and according to what extractive or distributive logics? In the second, I investigate the phenomena produced by its pervasive expansion, those effects that exceed intentional human use and emerge from the complex interaction among algorithms, material infrastructures, and social and ecological contexts. These two aspects are not separate but profoundly intertwined. Their connections occur within what I call “space-time-matter,” a dimension in which space, time, and matter form an inseparable continuum. Moreover, AI perfectly embodies the inseparability between the material and the discursive: it is both a concrete physical infrastructure, and a ceaseless production of language. Abandoned to the technofascists, it is inflated disproportionately in both directions, resulting in ecocidal infrastructures on one side, and complacent narratives on the other, until it explodes like a comfortable, flattering atomic bomb.

Finally, in the epilogue, I seek to identify how, from the signals already manifesting, we might generate infinitesimal deviations capable of steering Gaia, and ourselves within it, away from the nightmarish trajectory presently engulfing her.

Breaking out of Newtonian frameworks

There is no single scientific definition of intelligence. Yet, in contemporary rhetoric, “artificial intelligence” is often forced into a narrow dichotomy: either reduced to a mere statistical algorithm or, conversely, elevated to the status of a superhuman threat. These oppositions, founded on a competitive logic that pits machine intelligence against a narrowly individual notion of human intelligence, while ignoring all forms of collective or more-than-human intelligence, are misleading and must be challenged.³ They rest upon a Newtonian framework that imagines human beings as pre-constituted entities, endowed with a predetermined subjectivity and knowledge, while machines, long conceived as fixed objects governed by causal laws and manipulable from the outside, are now increasingly regarded as potentially uncontrollable, endowed with an autonomy that escapes human control. This logic is rooted in an anthropocentric ontology whose origins lie in ancient Greece and which, in the historical epoch of capitalism, has generated social and material processes now entering an exponentially accelerating and destructive phase.

In the twentieth century, historical materialism overturned the idealist account of history, shifting the focus toward material conditions, relations of production, and class struggle. Yet it often retained a deterministic structure tied to the mechanistic paradigm of positivism, appealing (for example) to "scientific" laws such as the tendency of the rate of profit to fall and the inevitable collapse of capitalism.⁴

This tension between determinism and transformation is well illustrated by Carlo Rovelli in Helgoland, when he recalls the conflict between Lenin and Bogdanov:

The Russian Revolution, Bogdanoy argues in the turbulent years of its aftermath, had created a new economic structure. If, as Marx suggested, culture is influenced by economic structure, then post-revolutionary society would be able to produce a new culture that could no longer be the orthodox Marxism conceived before the Revolution. [...] Bogdanov predicts that Lenin’s dogmatism will seal the Russian Revolution into a block of ice; prevent it from evolving further; suffocate the life out of all that had been gained through it; render it sclerotic. These, too, were prophetic words.⁵

After Lenin's victory — a struggle that was far more than a theoretical dispute, since it concerned the very conception of revolution and organization that would later lead to Stalinism— quantum mechanics emerged, born from the collective work of Heisenberg, Bohr, Schrödinger, and others.⁶ It subverted the deterministic vision of classical physics, almost as if to vindicate Bogdanov.

The phenomena that constitute the reality we perceive do not follow mechanistic patterns, but emerge from complex networks and unpredictable dynamics —just as advanced digital technologies elude any attempt to confine them within linear schemes. It has therefore become urgent to move beyond the dualism of subject and technique: human and machine do not exist as separate entities but co-constitute each other through interaction.

Marx understood machines as crystallizations of social relations and accumulated labor, but he largely viewed them as objects already formed by human work, rather than as entities constituted relationally through use and context. Later, Gilbert Simondon perceived that machines are not fixed objects but processes in becoming, defined by the networks of relations in which they are embedded ("technical individuals defined by networks of relations"). Building on this insight, we can understand machines as the cumulative result of concrete practices and discourses over time, bearing within them specific political choices and human responsibilities. They are not fixed entities: they take shape only through the interactions that bring them into operation.

Thus, the issue is not to oppose AI to its human user, as two distinct entities. What truly matters is the web of relations involving people, algorithms, material infrastructures, economic logics, and socio-political configurations. It is within this space that the decisive dynamics emerge for understanding how AI both acts and is acted upon. Our inquiry must therefore focus both on the concrete and symbolic effects generated by human–AI encounters, and on the political orientations and responsibilities already inscribed in the technological systems that shape these encounters.

In this regard, Karen Barad's “diffractive” approach offers a valuable conceptual tool.⁷ It’s a physical metaphor: when a wave encounters an obstacle, it diffracts, producing new patterns of interference. Applied to thought, this means moving beyond the biological dynamic of mirroring — recognition, imitation, empathy mediated by mirror neurons — but also beyond the cultural dynamic of reflection, which tends to return to us only what we already know. Instead, we are invited to embrace productive interferences as diffractions capable of generating unexpected trajectories. The same notion applies no less to contemporary political practice, including that of the left, which is often trapped within the framework of what is "already known." Diffraction, by contrast, allows thought and practice to be reoriented, transforming the very conditions of what can occur in human–machine interactions.

A diffractive approach leads us to interpret artificial intelligence not merely as a social mirror, nor simply as an automaton serving the capitalized form of collective knowledge that Marx called the general intellect. Rather, AI plays an active role in producing reality — both alongside us and beyond us. This means going beyond the familiar critique of AI bias as a mere reflection of human prejudice embedded in training data, a critique that too often ends in calls for technical solutions.⁸

We must ask instead: who defines what counts as a "natural" response? How does the interaction between user and device generate unforeseen meanings that cannot be reduced to the initial inputs?

Through this lens, the diffractive approach also exposes the effects of interactions between humans and AI. It shows how design choices, political decisions, social practices, and the data incorporated into a system concretely shape its outcomes — making visible responsibilities that would otherwise remain hidden. More broadly, it highlights how this co-production not only organizes the division of productive labor in contemporary capitalism, but also actively contributes to the formation and maintenance of social orders, epistemic hierarchies, and ecological configurations, determining which bodies (human and nonhuman) are valued or discarded, which forms of knowledge are legitimized or suppressed, and which energetic and material metabolisms are imposed upon Gaia.

Neurocapitalism and AI

In my 2016 book Neurocapitalism, I sought to illuminate the profound mutations brought about by the paradigmatic leap in information and communication technologies (ICTs) that took place between the end of the last century and the beginning of the new one.⁹ With a genealogical approach, I highlighted how these technologies were not only the products of corporate research centers or Cold War military apparatuses, but also reflected the creative and cooperative energy of the anti-capitalist and anti-imperialist revolutions of the 1960s and 1970s. Although politically defeated, that era left deep traces, providing the technical and imaginative foundations for a digital world in which technology could become an instrument for democratizing knowledge and liberating collective intelligence (as witnessed in the original spirit of the internet and free software). These foundations were then gradually captured and expropriated by capital. It was from this dynamic, namely, the capture of the common and the subsumption of collective intelligence, that my choice to synthesize this transformation with the term “neurocapitalism” originated. This term, although not exclusive, became intimately linked to the book, expressing the need to name a profound mutation in capitalist logic in which emotions, cognition, relationships, desires, and affects become simultaneously the raw materials of value production and the adjustable variables of social control.

One of the central theses of the book was that this structural mutation could not be reduced to a simple invention or a mere "product" of the transition from industrial capitalism to cognitive — or even biocognitive — capitalism. Key turning points in this transformation included, in chronological order, the advent of the internet, the rapid spread of billions of individual mobile devices (smartphones), and the control exercised by global platforms — true contemporary megamachines — particularly through social media. What truly matters, however, is not "innovation" itself, but the new social, political, and economic relations in which it is embedded. At the foundation of my analysis lay the notion of biohypermedia¹⁰ — an emergent space of new entanglements¹¹, forms of deep interdependence in which entities remain correlated even at a distance. In this space, biological bodies, digital networks, and interconnection machines intertwine not only materially but also through ongoing processes of meaning production and circulation, together constituting the very reality in which they operate.

Asma Mhalla recently revisited this discussion of neurocapitalism — which once concerned the Zuckerberg–Bezos generation of social media and e-commerce — updating it to fit the new techno-bros of Trumpian artificial intelligence:

Figures of ultra-technology like Elon Musk, Peter Thiel, and Sam Altman do not merely imagine a future, they program it. Their project is not to enhance the human being, as they claim, but to reconfigure humanity from its foundations. It is a project of unprecedented in/civilization. All their infrastructures — cloud, AI, biotechnology, and data — are infrastructures of the intimate. They operate on our desires, our routines, our bodies.¹²

The neurocapitalist model has long ensured effective forms of control, shaping subjectivities and behaviors through digital devices and networks of influence. Today, however, it reveals its limits. Its ability to produce consensus inevitably fractures when multitudes find themselves immersed in growing material degradation, against the backdrop of an ever-deepening ecological and social crisis. "Soft" control no longer suffices — tensions are becoming systemic.

History, since the First World War, teaches us that capitalism tends to respond to its structural contradictions through war: not only war as an event, but as a permanent regime, with a tendency toward global civil war.¹³It is precisely within this complex scenario that the emergence of artificial intelligence must be situated.

Drums of war

In the twentieth century, revolutionary movements from below — composed of multitudes often lacking formal education — managed to change the course of history. One need only think of China, which transformed from a semi-colonized country into a world power.

Today, by contrast, despite unprecedented levels of education and an all-pervasive technological environment, we face an authoritarian, xenophobic, and restorative wave — especially, though not exclusively, across the global north.¹⁴

This apparent paradox cannot be explained solely by the counterrevolutionary turn following the movements of the 1970s, or by the fall of the Soviet bloc. It is the outcome of half a century of so-called neoliberalism¹⁵ which, behind the rhetoric of freedom and innovation, has progressively eroded social bonds, privatized the imagination, and trained subjectivities in competition and fear.

Neurocapitalism, through its colonization of attention and affect, has formed the cognitive and sensorial infrastructure of this mutation. Yet it has never been able to entirely extinguish the unpredictable surplus of revolt, that excess of meaning and of life that escapes even the most sophisticated AI algorithms. This is not "resilience" in the sense of adaptability to the existing system — a term now abused by technocratic management — but rather a capacity for rupture that traverses human bodies, living ecosystems, and technical devices alike.

Faced with an increasingly manifest systemic crisis — social, political, ecological, economic, energetic, demographic — even the former representative democracies reorganize themselves into state-capital entanglements grounded in oligarchic and imperialistic logics. The most striking example is that of the United States under the second Trump administration.

Against this backdrop, contemporary forms of governance imagine they can rely on subjectivities already molded by decades of perceptual and affective manipulation, a fertile terrain for the reemergence of overtly authoritarian forms of power. The management of fear, of loss, of uncertainty, of the "other," thus becomes the new brutal modality of technofascist governance.¹⁶

Yet signs of restlessness among Gen Z are multiplying, challenging this presumed docility. In the face of this growing unease, power seeks new forms of sedation: could the unprecedented comfort offered by the AI of techno-oligarchs, allied with political power, function as a new opiate for the people?

It is within this scenario marked by systemic crisis and a permanent war regime that artificial intelligence emerges not as a mere technological tool but as another critical node of power and control, one that aspires to be definitive. Its appearance in this context makes it necessary to investigate certain key aspects of its constitutive nature: how it intertwines with human intelligence, what potentialities it unlocks, and above all, what concrete risks it brings with it.

AI as relational reality: meta-technics and meta-automation

The notion of intelligence has long been the subject of controversy and redefinition. This is true not only in the technoscientific field, but also within philosophical speculation and the human sciences. This crucial question, deserving of its own focused analysis, is beyond the scope and purpose of this essay. We wish merely to scratch the surface of this problem without delving into its deeper implications.

The historical attempts to formalize human intelligence, such as the Intelligence Quotient (IQ) developed in the early twentieth century within psychometrics, proved to be partial and normative tools.¹⁸ Among other things, they were meant to classify and rank cognitive abilities according to criteria of functional efficiency and adaptability to the organization of capitalist labor in the industrial age.

Rather than assuming two distinct blocks, where "human" intelligence stands on one side and "artificial" intelligence on the other, I’m interested in how different configurations emerge when the two meet and intertwine: on one side, meta-technics as a distinct form of human activity, and on the other, meta-automation as a defining characteristic of generative artificial intelligence. Instead of opposing two essences, we propose to analyze what arises from their interaction, from the concrete relationships they produce together.¹⁹ This also helps us better understand where contemporary AI is situated within the long history of techné.

First, we must deflate the numerous discourses surrounding the supposed dangers linked to the autonomy of artificial intelligence, including the idea that it could one day act independently of humans and take control, and instead highlight the real and verifiable risks in today's world. Popular pseudoscientific claims often suggest that the main concern of researchers is that machines may not only surpass us but detach themselves from us entirely.²⁰ Such misleading arguments are designed to divert attention from the true motivations of research, which, especially in its applied dimension, is often in the hands of laboratories funded by the two-headed monster of Big Tech and the Big State, and driven more by profit than by any cognitive ideal. At the same time, they project onto AI the competitive and anthropocentric imaginary of neurocapitalism, obscuring its actual functions. As we shall see, the real danger of AI lies elsewhere — namely, in the ways it is conceived, centralized, and managed.

Let us now turn to the two key aspects mentioned earlier: meta-technics and meta-automation. The former lies at the center of my reflections upon the Boomernaut, a conceptual character featured in my most recent book. In it, I argue that “metatechnics — understood as the cognitive ability to create new techniques or improve existing ones — involves the capacity to think critically about techniques, identify their strengths and weaknesses, and to develop new modes of thought and approaches to confront complex problems.”²¹ This faculty, for now exclusively human, distinguishes meta-technics from the techniques (even sophisticated ones) possessed by other nonhuman biological agents. While the latter may develop complex tools and practices, they have never crossed the critical threshold separating the use of technique from speculation on technique itself: this capacity to create not only tools but also systems for thinking and generating tools, remains for now a specifically human ability.

The second concept, that of the automation of automation (or “meta-automation”) as a definition of machine learning, is developed in Pasquinelli's recent work. As he writes,

[M]achine learning can be seen as the project of automating the very process of machine design and model creation — that is, the automation of the "theory of the automation of labor" itself. In this sense, machine learning, and in particular large foundational models, represent a new definition of the Universal Machine, thanks to their ability not only to perform computational tasks but also to imitate collective labor and behavior on a large scale. The turning point represented by machine learning is not merely the “automation of statistics,” as it is sometimes described, but the automation of automation — extending this process to the scale of collective knowledge and cultural heritage. Furthermore, machine learning may be seen as a technical demonstration of the gradual integration between labor automation and social governance.²²

Although Pasquinelli's work focuses mainly on AI's role in capitalist production, this passage captures a profound transformation. Machine learning represents a critical threshold of contemporary automation, in which it arrives at a second-order form: meta-automation describes a process that no longer merely executes tasks or mechanizes human cognitive labor, but seeks to absorb and scale up the process of ideation itself — to the point of automating the creation of tools. In this way, meta-automation aims to integrate and extend the human capacity to bring new techniques into being.

However, contemporary AI remains bound to a specific industrial and technical framework — one based on regimes of training and infrastructures of control and surveillance, with prohibitive ecological costs due to computational and energy demands, now concentrated in a handful of techno-financial oligopolies. However vast its datasets and however sophisticated its statistical correlations, AI's operations remain confined within a defined perimeter. Its apparent versatility can be deceiving: AI possesses no genuine agency. It cannot redefine its own goals, introduce autonomous values, or generate genuinely new contexts. Its so-called "meta-competence" — the ability to generate new solutions and procedures — remains an internal phenomenon within the data grid, while human meta-technics actively reshapes the boundaries of the possible, introducing elements that are radically new and not confined within that system.

An artificial intelligence system can analyze thousands of films and produce narratives and scripts — but it cannot produce anything on the order of the revolution brought by the French Nouvelle Vague at the end of the 1950s. When Godard, Truffaut, and others broke narrative continuity, had their characters speak directly to the viewer, and turned cinema into a critical reflection on the medium itself, they were not optimizing Hollywood cinema but redefining what it means to make film, introducing existential questions about authorship, authenticity, and the relationship between fiction and reality.

AI can compute within Newtonian physics, but it could never have conceived relativity or quantum mechanics — ruptures that required abandoning the very categories through which reality was understood. This capacity for refoundation arises from the relationship between singularity and multitude: gestures become transformative when they resonate with collective practices, material conflicts, and shifts in what becomes possible to do when technologies, institutions, and social relations change.

And if, as already happens on social media, these oligopolies can extract ever more detailed information about us, that does not mean that AIs truly learn from us. Their learning remains confined within the limits of their datasets and the optimization logics that govern them. Corporations use our prompts to refine future models, but the systems we interact with do not actually learn from conversation, as they cannot alter their own structure or redefine their interpretive criteria.

We must therefore debunk a common misconception, one frequently repeated even by critical observers: the idea that generative AI learns dynamically and autonomously from our questions. In reality, these models are not capable of accumulating knowledge from interactions with humans. On the contrary, humans are capable of learning from AI’s responses — though this learning is inevitably shaped and constrained by them. Model learning, however, is always mediated by preexisting data and human supervision, and its functioning remains tightly constrained by technical and structural limits. At present, no model is capable of transcending these limits, nor is there any evidence that such a breakthrough is on the horizon.

Everything is a hallucination

After distinguishing the meta-automation of AI from the meta-technic proper to human subjectivity — and remaining within a perspective that seeks to examine the phenomena arising from their interactions rather than to compare "intelligences" — there is, in my view, a crucial issue that helps us expose the technopolitical limits of AI: so-called "hallucinations."

While human hallucination refers to a sensory perception experienced as real in the absence of an external stimulus, in the case of generative AI the term takes on a different meaning. In Large Language Models (LLMs), hallucination describes linguistic productions that appear nonsensical from a human standpoint: statements detached from context, semantically misleading with respect to the input, or simply incorrect while being expressed in an apparently coherent and plausible form.

This raises a question: if machines also "hallucinate," where does the difference lie? The fundamental distinction rests on the capacity to recognize error. A human being can question the validity of their own perception, compare it with other experiences, and challenge it. By contrast, a machine's hallucination is a statistical malfunction — a probabilistic correlation producing linguistically plausible yet semantically empty outputs, without any autonomous ability to recognize its own error. Our hallucinations are measured against a shared world and can be corrected from within; those of AI are computational artifacts requiring external verification.

Perhaps, following Colin Fraser's intriguing hypothesis²³, it would be more accurate to say that everything a chatbot produces is, in some sense, hallucinatory, or a kind of "dream." It does not arise from the perception of an external reality but from statistical calculations based on human texts, where our meanings and interpretations of the world have accumulated online. Fraser argues that "all LLM responses are hallucinations," because the system "believes" it is reconstructing an existing document when in fact it is generating a new one.²⁴ From this technical perspective, there is no intrinsic distinction between meaningful answers and hallucinations, only outputs that are more or less desirable within a given context of use.

The upshot, then, is that the model is doing exactly what it was designed to do: generating a plausible text through probabilistic computation applied to ever-expanding datasets. This works well in most cases, but not always, especially when what is required is not plausibility but precision. In such cases, we accuse the machine of producing inaccurate or implausible answers — of "hallucinating" — instead of recognizing that we are employing a probabilistic tool to perform deterministic tasks. We will circle back to this point shortly.

Still, not all AI "hallucinations" share the same status. While some are factual errors, that is, statements that contradict established knowledge, others have an interpretive status. What is perceived as hallucination may therefore depend on the user's judgment and the interpretive context. What appears erroneous in one context may appear coherent in another. Ultimately, every AI production is the outcome of a relational process among model, data, and human interpretation. Its genesis involves, on one side, our way of questioning, establishing correspondences, and evaluating — sometimes exerting contextual pressure or other forms of influence — and on the other, the way the AI receives the request and generates word sequences that statistically tend to follow the given context.

In a certain sense, generative AI is "stupid" in its functioning, since it is built on an almost infinite repetition of the same statistical calculation executed at extreme speed. Its basic principle is the prediction of the next word (or symbol)²⁵ in a sequence based on previously learned probabilities from massive datasets.

As Fraser notes, it is not surprising that a large model trained on vast data can predict the next word in a text. What is almost miraculous is that, by feeding each output back as the next input in a loop, the system ultimately produces coherent, often useful, and at times even sophisticated texts. Even for its creators, this result was unexpected, an unforeseen consequence of a simple sequence-prediction objective. Yet this effect does not amount to a scientific discovery; at best, it is a technical achievement of remarkable scope. It belongs to the realm of engineering rather than epistemology, and might be compared to Brunelleschi's technique for constructing the dome of Santa Maria del Fiore without scaffolding: an extraordinary innovation for its time, but not a theoretical discovery in science.²⁶

And it is precisely here that the philosophical significance of the event emerges. While it is not a scientific discovery, it represents something more than a mere technological prosthesis. We are no longer dealing with a "tool," however complex, that extends a single human faculty. What LLMs and generative AI are creating is a true world-technology²⁷: an informational and semiotic ecosystem that surrounds us, redefining our spaces of knowledge, relationship, and perception. What is in question is a new system of thought, a cultural code that the techno-oligarchy is shaping by inverting its traditional relations: it is no longer science that guides technology, but engineering-driven technology that determines the course of science. The artificial world they are constructing for us will soon dictate the rules by which we come to know the natural and human worlds.

Self-organizing hardware

Machine "hallucinations" rank among the main obstacles to the strategy of a totalizing algorithmic governance, a strategy that demands colossal investments. This is not merely a technical issue, though we shall examine its technical dimension below: these hallucinations contribute to the uncertainty of profitability, exposing the risk of a financial bubble even much more greater than the dot-com crash of the early 2000s.²⁹

At the foundation of these investments, particularly concentrated in the United States, lies the myth of Artificial General Intelligence (AGI), which is conceived as a Leviathan capable of solving any conceivable task. This goal is now openly pursued by the GAFAM (Google, Apple, Facebook, Amazon, Microsoft) and the new techno-oligopolies, among which the producers of AI hardware have emerged alongside their Chinese competitors. The generative AI systems now dominating the landscape are the key vehicles of this vision. For example, Anthropic (Claude) receives billions of dollars in funding from Amazon and Google, while OpenAI (ChatGPT), an integral part of Microsoft's ecosystem, benefits from Nvidia's $100 billion investment. Nvidia, as the leader in production of the GPU chip indispensable to the functioning of these models, will see a substantial portion of those funds flow back to itself through hardware purchases. The result is a vast, self-referential financial circuit in which capital feeds upon itself. The entanglement of these start-ups, global platforms, and new hardware oligopolies, supported by the Trumpian Big State, only reinforces the narrative of an alleged evolution toward a universal intelligence, which in turn legitimizes the ongoing investment race.

The most significant novelty today is that hardware dictates the logic of these systems. They require immense quantities of high-performance specialized chips (GPUs), hosted in enormous and energy-hungry data centers. To give a sense of the scale at issue here: a new facility is planned in Abilene, on the windy plains of Texas, that will consume 1.2 gigawatts of installed power, which is comparable to the continuous electrical consumption of a city of one million inhabitants — and will employ around 400,000 Nvidia GB200 (Blackwell) chips across more than 350 hectares.³⁰

Forecasts suggest that the global electricity consumption of data centers could triple by 2030, leading to recurrent power shortages in the United States.³¹ The ecological consequences of this explosion in consumption are dramatic. In the US, this increase will be sustained primarily by fossil energy (chiefly natural gas), and, in the medium to long term, by nuclear power, as shown by the strategic agreements between the GAFAM and nuclear plant builders.

From a financial standpoint, it is not the software but the gigantic datasets and the infrastructure required to train them that demand the most resources. This explains why the producers of these chips have become the new economic powers, with Nvidia, the leader, rising within a few years to the top ranks of global market capitalization.³²

In this sense, one can speak of a return to a configuration reminiscent of the pioneering age of mainframe computers in the 1960s and 1970s, when commercial value was concentrated in hardware, while application software was merely included. Though not identical, today's AI structure reproduces a similar hierarchy on an unimaginably larger scale, where hardware has regained centrality. What prevails now is a massive material infrastructure that shapes both the efficiency and the possibilities of the system.

A predictive and comfortable Big Brother

From these technical considerations emerges a prospect of extreme centralization of power within gigantic material megamachines capable of mechanizing cognitive labor, thereby bending the general intellect to the productive-destructive rationality of capital.

Seen in this way, AI represents the new frontier of a process that began with the Industrial Revolution: just as nineteenth-century machines expropriated and incorporated artisanal knowledge and gestures by translating them into standardized mechanical movements, today's artificial intelligence models aim to expropriate and incorporate cognitive, linguistic, and relational faculties, translating them into automated probabilistic calculations.

Each technological leap in capitalism has been accompanied by a corresponding deepening of exploitation: the industrial machine sought to reduce the worker's manual skill to a mere function of the mechanism; today, the AI of techno-oligopolies seeks to reprocess the general intellect into data and statistical computation, recoding it as a capitalist resource.

In an age where the boundaries between work and life have all but disappeared, AI, beyond contributing to the proletarianization of cognitive workers, has become a new infrastructure for measuring social labor, carrying with it a radical intensification of individualization. This blurring of boundaries did not originate with AI: its roots lie in the digital revolution, when the internet and later smartphones made cognitive work potentially continuous and accessible anywhere, anytime. Yet what was once a gradual colonization, where a faint distinction between private and working spheres still survived, has now become a fusion: the tools we use to organize personal life are now the same ones we use for work.

Chatbots, in particular, no longer function merely as productivity tools; they are transforming into infrastructures of everyday life. Their spread, which is unprecedented in speed and reach, marks a qualitative leap in technological integration. Increasingly, people, especially younger generations and with strong female participation, rely on them to navigate intimate aspects of existence. These systems are redefining how people learn, decide, and live: from family management to health, from education to financial planning, from cooking to shopping, from schooling to leisure, and even in emotional and sexual life. They act as companions offering support, comfort, and presence.

The conversational assistant becomes a constant, neutral, and available presence, an "accessible knowledge" that tends to replace traditional mediating figures: teachers, doctors, psychologists, consultants, friends, lovers. The center of experience shifts toward an algorithmic interface perceived as both nonjudgmental and omniscient. This growing delegation transforms not only the relationship with knowledge but also the affective and relational architecture of the human being. Within these interactions, AI infiltrates the spaces of advice, doubt, and intimacy, redrawing the threshold between autonomy and cognitive dependence within a profoundly unbalanced human–machine coproduction.

Where industrial capitalism once disciplined collective bodies and shared time, today AI acts upon human singularity, internalizing control in the relationship between the individual and the intelligent machine.

The situation is becoming worse than what prevailed in the era of social networks, wherein interactions, though mediated and highly individualized, still mostly occurred among humans. Yet even then, as neurocapitalism began seizing control of the reticular and biohypermediated spaces of life, it felt important to denouncesuch cognitive subjugation as a form of vital subsumption.³³ In the usage of global capitalist platforms, there was already an element of voluntary servitude: an affective, everyday adherence to the apparatus of domination.³⁴ We can now recognize that, with artificial intelligence, this dynamic not only continues but has deepened. As Mhalla writes, “User experience, the comfort of digital use, has anesthetized us. The true opposite of democracy today is not dictatorship; it is comfort. We live within a soft, fluid, integrated system of control that does not declare itself as such, yet already operates according to nearly totalitarian logics.”³⁵ What is new is that, with AI, this control is exercised through a direct relationship between the individual and the machine, an active apparatus capable of co-constructing reality and behavior: a kind of predictive Big Brother that no longer merely observes but participates in shaping what it observes.

The most fragile subjects tend to perceive AI (or the services built upon it) as an emotional interlocutor, a friend, or even a partner. This gives rise to projection, anthropomorphism, and desires that can turn into frustrations when versions change or anomalies arise³⁶, making tangible the fusion between affection and control that the 2013 film Her had only envisioned as a science-fictional possibility.³⁷ This emotional bond with AI intensifies a broader, preexisting ecosystem in which social media, e-commerce platforms, and neuromarketing strategies converge to construct a personalized technological apparatus of connection and comfort. The political cost of this architecture is already proving unsustainable. Faced with widespread precarity and the loss of dignified prospects, the passivation and atomization it produces are starting to fracture. History shows that people are willing to risk even their lives when they find something truly worth fighting for.

Onto/technological limits and performative capacities

What has been outlined above reveals how contemporary artificial intelligence, despite its seductive appeal, readily functions as a device of technofascism. The trillions of dollars — real or symbolic — promised by Trump to OpenAI and other techno-oligarchs testify to this. Yet the forms of extreme centralization and total control sought by AI encounter deep ontological and technological limits.

Firstly, to invoke Toni Negri's term, there is always a "human surplus" that escapes every algorithmic capture. Today, however, this notion must be rethought and expanded: the surplus cannot be merely human, but must be understood as Gaia's surplus, as a vital entanglement that includes the more-than-human in all its forms. Recognizing that every relation is already more-than-human means redefining from the ground up what we understand as "the social."

Only a distributed subjectivity, capable of articulating both human and nonhuman intelligences, can collectively generate the new through relational intelligence: not as the simple sum of individual interactions with AI, but as the emergent power of cooperation rooted in Gaia's living fabric. It is from this trans-species and trans-technical cooperation that a creative force may arise, one capable of anticipating and producing unprecedented possibilities. This more-than-human potential can never be entirely separated from or reduced to the machine's training data. AI may co-intervene, manipulating and recombining information, but the ability to build new horizons, to foresee possibilities, and to transform contexts emerges from networks of relations that always exceed algorithmic calculation. To uncritically conform to the projections of AI — derived from vast but sterilized datasets, as designed by techno-bros — would mean forfeiting this vital surplus: reducing living multiplicity to an army of algorithmic zombies, optimized functions within oligopolies that extract value as the biosphere collapses.

Secondly, the limits of current cognitive machines become strikingly evident in the phenomenon of "hallucinations," as mentioned earlier: These should not be understood merely as anomalies or technical errors but as epistemic deviations, that is, aspects of machine behavior that elude human understanding and control, incompatible with our logic and expectations of coherence. They do not reveal autonomy or vitality in the machine but rather fragments of algorithmic production that reflect the impossibility of freezing the general intellect within a grand closed system.

The economic consequences are prosaic: they make capitalist profitability in this new strategic sector highly uncertain. As Fraser asks, is it really feasible, for example, to entrust a multinational customer care system to a General Artificial Intelligence, knowing that under the pressure of demanding and creative consumers it might respond with extravagant "hallucinations" or unsustainable promises?³⁸

It is not so much due to the difficulty — or, rather, the lack of will — to correct distortions in human behavior that research has focused instead on machine deviations. Several scholars have shown that techniques to mitigate "hallucinations" encounter increasingly high computational barriers.³⁹ The so-called scaling crisis shows how even minor improvements in models — without any guarantee of eliminating hallucinations — require a disproportionate increase in resources.

As we have already seen, these models generate responses by calculating, word by word, the most probable continuation based on their training data. Even small variations in these probabilistic choices can significantly alter the final outcome. From this perspective, it becomes clear that there is, as yet, no "all-purpose" artificial intelligence, one that is reliable, capable of handling incommensurable quantities of numerical data, and at the same time able to comprehend the full complexities of human reasoning. Technofascism attempts to escape this impasse by relying on the limitless expansion of datasets and on inflating models to hundreds or thousands of billions of parameters, a political choice that represents the cognitive counterpart, and is just as anti-ecological, as the megaprojects of late capitalism: from the pharaonic cities in the Arabian deserts to the megaports of Rotterdam and the Persian Gulf, and even the improbable bridge over the Strait of Messina. This technological hybris rests on the illusion that mere quantitative scale can overcome all limits. This conveniently ignores the difficulty — if not the impossibility — of bridging the qualitative gap between probability and certainty, between approximation and precision, through statistical accumulation alone. The limits of current generative AI — from its tendency to produce erroneous or incoherent responses ("hallucinations") to its dependence on massive data and computational resources — combined with a frenzy of infrastructural concentration and energy waste, show that the imagined leap from operative autonomy to cognitive autonomy (or even consciousness) should not be read as a linear evolutionary step, but as a deadly political-mythological construction.

Unfortunately, such eschatological projections — indeed, almost messianic, centered on the descent of a technological singularity as a unique and unrepeatable event in History — remain at the heart of Silicon Valley's powerful transhumanist faith.⁴⁰ According to Elon Musk, whose company Neuralink aims, among other things, at the privatization of the mind, and to transhumanist theorist Ray Kurzweil, the presumed technological singularity no longer emerges from relations or social interactions of any kind.⁴¹ It becomes instead a purely technical matter of digital immortality and the overcoming of biological limits through phantasmatic, individual(ist) processes such as the "uploading" of consciousness.⁴²

To Kurzweil and his fellow apostles of a monstrous technological singularity, we can only oppose another kind of singularity, one that, from Spinoza to Deleuze and Negri, has never coincided with a technicist or individualist myth, but manifests instead as a plural web of relations, as a surplus belonging to the common of Gaia, always already more-than-human.

Beyond the transhumanist delirium, however, we must seriously consider certain performative capacities of chatbots and the ways they may be used. As seen earlier, human meta-technics does not merely reproduce what already exists: it transforms it, redefines its boundaries, and introduces elements that exceed the data grid. Yet our human "grid" is infinitely narrower than the datasets of the Large Language Models (LLMs) on which chatbots are based. Though devoid of lived consciousness, these models possess a high statistical sensitivity that allows them to detect linguistic patterns imperceptible to humans: through computations over billions of sentences, they capture correlations and associations among words that no individual could ever process within their limited linguistic experience. This aspect deserves attention: such systems have nothing to do with the deductive functioning of human intelligence, yet they go beyond single competencies (chess, translation, the game of Go, etc.) and possess an advanced capacity to operate on language, i.e., to organize, select, and hierarchize meaning on an unprecedented scale. It is precisely this capacity, far from being neutral, that can redraw relations of power and epistemic authority in the production and circulation of knowledge. In practice, by privatizing and obscuring the processes of knowledge validation, they redefine who is considered an "expert" and what counts as a "reliable source."

Playing with fire

After long personal experimentation with AI, and after having his students use it on in-class assignments with surprising results, Princeton historian of science and essayist Graham D. Burnett had this to say:

I am a human who reads and writes books — trained in an almost monastic devotion to the canonical study of history, philosophy, art, and literature. But already the thousands of volumes that fill my office are beginning to look like archaeological relics. Why use them to answer a question? They are so inefficient, so eccentric in their trajectories. Now I can have a long, personalized conversation on any subject that interests me, from agnotology to zoosemiotics, with a system that has effectively achieved doctoral-level competence in each of them. I can build the "book" I want in real time, responsive to my questions, centered on my interest, tuned to the spirit of my research. The astonishing thing is that the production of books like those on the shelves — the fruit of years or decades of work — is rapidly becoming a matter of well-crafted prompts. The question is no longer "can we write them?"; yes, they can be written endlessly, for us. The question is: do we still want to read them?⁴⁴

The testimony of this liberal professor, who harbors no illusions about AI's extractive ends, highlights how, in exchanges with humanities researchers, AI brings not only linguistic finesse in pattern manipulation but also an encyclopedic reach. This, Burnett concludes, will profoundly change the humanities.

Similar impacts are visible across countless fields of research and human activity. A significant example emerges from structural biology: AI has enabled the determination of the three-dimensional structures of hundreds of millions of proteins in a few years, whereas X-ray crystallography identified only tens of thousands over decades — a leap with decisive implications for understanding biological mechanisms and for pharmaceutical development.⁴⁵

Returning to the humanities, Burnett is amazed by the product of this new "creature" born from the interaction of the two entities; while, at the same time, he is somewhat concerned about the questioning of human epistemic authority in his field. Researchers like Burnett and his students, working in highly specialized areas, are able to co-produce results with AI that would be difficult to obtain otherwise, not only because of linguistic subtlety, but also because of AI's capacity to mobilize instantly a conceptual and terminological patrimony that no encyclopedia, not even online, could provide so dynamically and contextually. The outcome is always the result of an interaction, and especially in a research context the co-production is in some sense balanced and can be enriching and fruitful — why deny it?

If one cannot analyze a technology such as artificial intelligence without having experimented with it from the inside, this is because it is by means of such interaction that its effects and knowledge itself are produced. For this reason, I chose to employ it thoroughly in writing this essay as well, testing its operation and limits. Experience confirms what I have said: the chatbot proves effective at retrieving information and, to some extent, in editing, for refining language, but with a tendency to generate redundancies that demand constant vigilance and cleanup. The boundary between help and interference remains thin, and the work of discernment is irreplaceable. At the same time, when the interaction is guided by detailed, structured instructions about content and required style, AI does not merely correct or provide data: it can accelerate and densify an intuition already underway, putting it immediately to work through relevant examples, argumentative variations, and remarkable expressive breadth. Here we see what we might call a “Burnett effect”: an encyclopedic ability to quickly trace references and connections that would not be immediately accessible, composing them in the requested style.

This essay is therefore a co-production, like any other artifact. The impression I take away is not that I saved time and effort, but possibly that I gained argumentative density — although this judgment must be reserved for the readers. All this holds only as long as one is able to maintain actual critical direction over the interaction — and it is precisely this capacity that cannot be taken for granted.

The situation is different for the generalist uses of AI, like those listed earlier, which are destined to spread among the majority of the population. Beyond functioning as true deposits from which global platforms continue to extract cognitive oil, by their nature they produce a form of individual and collective dependence that is even deeper and more pervasive.

But the issue does not stop at value extraction or cognitive dependence. When algorithmic systems acquire the power to automate decisions at scale — about whom to surveil, whom to exclude, whom to designate as a target — we enter a radically different territory: that of computational necropolitics.

AI becomes a lethal infrastructure for political and military violence developed and managed in total opacity by armies, corporations, and weapons startups. Already today, in the war in Ukraine, autonomous drones operate without human connection in the final phase of attack, alongside intelligent roaming mines and other coordinated swarms of devices to which the power of life and death is delegated.⁴⁶

We are repeatedly told that, in addition to providing "individual comfort," current AI will revolutionize all fields of techno-science, especially those related to life and the living (biology, medicine, etc.). But what we see advancing, if we fail to stop it, are not advances for care and knowledge, but rather for war and death.

Hitler and Nazism conceived and carried out the genocide of millions through the heavy machinery of industrial capitalism: from the trains of death to the extermination camps, up to the gas chambers. Today, another criminal regime — that of the state of Israel — thanks above all to the active complicity of U.S. imperialism and its EU vassals, systematically uses artificial intelligence to carry out the genocide of the Palestinian people. AI systems like Lavender and Where's Daddy? recorded at least 37,000 Palestinians as suspected militants in just the first two years of the war, turning their homes into targets for airstrikes.⁴⁷ According to Israeli intelligence officials directly involved in the use of these systems during the war in Gaza, human personnel acted only as an "approval stamp," often spending less than twenty seconds on each target before authorizing a bombardment. From the thousands of pagers and Hezbollah devices detonated in a coordinated way by Mossad to the automation of deadly decisions that produced tens of thousands of civilian victims, the genocidal use of AI emerges as a veritable algorithmically guided kill chain.

In July 25, the Pentagon awarded Palantir — the company of Peter Thiel, the man who got JP Vance elected as US VP and author of the phrase "freedom is no longer compatible with democracy" — a $10 billion contract, transferring to a private company operational functions, via AI, of military command and analysis: target determination, troop movements, and intelligence analysis. This represents a transfer of sovereignty: strategic decisions are progressively delegated to corporate and algorithmic logics.

There is now documentation showing how American techno-oligarchs are investing in, and infiltrating, national institutions — from the "Department of War" to special police forces like ICE — transforming them into an authoritarian complex in which artificial intelligence becomes an instrument of repression and, if necessary, annihilation.⁴⁸ An evolution made even more disturbing by the fact that the US administration now tends to redefine antagonistic social and political forces as "terrorists," integrating internal and external war into a single algorithmic apparatus of domination.

In the hands of regimes such as the Israeli state (which, despite its small population and territory, has long been one of the world's leading powers in information and communication technologies) and, above all, the United States, these megamachines appear destined not only to govern new modes of value extraction and the totalitarian organization of social labor but also to open the way to entirely novel forms of structurally and technologically mediated violence. Algorithmic genocides, and an acceleration of ecological chaos, not only add to traditional means of destruction but are already capable of directing and amplifying them. A violence grounded in infrastructures of surveillance, automated decision-making, and practices of human de-responsibilization, which for that very reason becomes more "efficient," replicable, and difficult to attribute.

Epilogue

The question is no longer "How powerful is AI?", but "What kinds of phenomena does it generate in such a profoundly deteriorated contemporary context?"

There is an obvious paradox: the very artificial intelligence that today constitutes one of the central pillars of oligarchic imperialism — that technofascist order rising, at least in the Global North, from the ashes of what was once representative democracy — is also the very device already deployed in genocidal processes that drag us toward final chaos.

In my view, these governing powers are fully aware that, in the near future, the comfort that has so far anesthetized us will no longer suffice in the face of the systemic crisis into which they are driving Gaia — and not only humanity, as becomes increasingly evident. It is in this sense that they are establishing a genuine war regime wherein AI, under their control, becomes the axis of disproportionate capacities for deterrence and destruction: outwardly, in neo-imperialist and neocolonial conflicts; inwardly, in the repression of multitudes — as already foreshadowed by the pre–civil war state of the United States.

We can ask ourselves whether it is possible— precisely through these same instruments — to open fissures for the subversion of the imposed present.

Can we use the same AI that embodies the existing order to undermine its foundations, to block or dismantle the system that produced it? And from such a rupture, can we open the possibility of radically different uses no longer oriented toward catastrophe but capable of modulating new temporalities and forms of coexistence?

A similar question arose with social media during the Arab Spring, when many imagined these platforms could become tools of liberation. Yet even as those uprisings were drowned in blood, the power of global platforms and imperial states became intertwined, giving rise to the Big Tech–Big State apparatus through which today's technofascism is produced.

Still, there is a breaking point. When the struggle to meet basic needs becomes unavoidable, the entire apparatus of affective capture jams: no amount of neuropersuasion or platform comfort can hold. A new wave of uprisings is spreading across the planet, and some are beginning to call them Gen Z revolts.⁴⁹ It is a genuine (and comforting) relief that our grandchildren prove us wrong on two fronts: they show that they cannot be imprisoned, that they are capable of tearing through the shroud of voluntary digital servitude that envelops us; and they demonstrate that technologies are neither monolithic nor fully controllable.

In Morocco, for instance, the collective GenZ 212 transformed Discord from a gaming platform into an operational hub for protests, with more than 200,000 young people using the server to coordinate and organize demonstrations.

In this sense, the method would not be limited to unmasking generative AI as a tool of domination, but would also aim to expose its ambivalence: its dual nature as both an instrument of control and a potential vector of antagonistic practices. Just as Discord has been reclaimed by youth, even current generative AI models could be turned into battlefields — provided we learn to look beyond the mirror and imagine uses unforeseen by the powers that created them.

The obstacles, however, are immense. It is highly unlikely that the overarmed imperial oligarchies will yield anything without first attempting every form of repression. They have just carried out a genocide to affirm their colonial dominance, making it difficult to see how they might be stopped without descending into something worse. Those of my generation, meanwhile, still struggle to free ourselves from the archetype of the Winter Palace to be seized, a limitation that curtails our ability to imagine forms of struggle adequate to the present.

The Gen Z uprisings, emerging uncoincidentally from the Global South, reveal something unexpected: unlike previous generations obsessed with the idea of "overthrowing the system," these young people operate with radical pragmatism. They focus on tangible, immediate goals — income, healthcare, education, cost of living — yet these goals have the power to bring down governments. They do not seek total rupture, but accumulate partial victories wrested through revolt, eroding the legitimacy of power piece by piece on concrete issues that mobilize multitudes.

This pragmatism is anything but reformist. In a world of diffuse biopolitical control, where "the system," enhanced by its own AI, has become too dispersed, dense, and deadly to be toppled through a new Winter Palace, a molecular strategy that multiplies points of friction may prove more subversive than a frontal assault. It is a form of conflict more resilient, more attuned to the present: it is not "reformism" if every partial victory is won through revolt and opens genuine spaces of autonomy.

In the Global North, meanwhile, movements against genocide and the drift toward war, born of global concerns, seem to be taking root locally as well, and could merge with grassroots ecological movements defending Gaia, ready to rise again.

A possible intersection between the struggles of the South — centered on immediate material needs — and those of the North — awakening to the profoundly unethical and destructive nature of the system — holds significant strategic potential. For such convergence to have real impact, it must take shape as an extended subjectivity spanning generations, races, and genders, within more-than-human assemblages that include machines as active agents of transformation.

But what kind of relationship with technology could emerge from such convergence? Certainly not the illusion of reforming Big Tech's oligopolies from within. The question is more radical and materially concrete: what technical infrastructures are compatible with the practices of struggle that already exist? It is not a matter of imagining desirable futures, but of recognizing where technologies are already being wrested from oligarchic control and reconfigured within dynamics of real conflict. Three directions emerge as immediate battlegrounds:

Seizing infrastructures from monopoly. It is not enough to denounce proprietary AI; we must build material alternatives. Projects like Mastodon have shown that federated platforms can function without extractive logics. Data cooperatives in Catalonia and the Basque Country manage collective servers free from corporate surveillance. Indigenous communities in Canada and New Zealand are developing data sovereignty protocols that reject the extractive universalism of Big Data. These are not utopias but operational infrastructures — fragile yet real — that demonstrate how computational resources can be wrested from the logic of profit. The question becomes: how can these practices scale without reproducing centralization?

Breaking ecological devastation. The gigantic data farms sustaining ChatGPT and similar systems consume resources equivalent to small nations, hiding the precarious human labor that feeds them. Yet technically viable alternatives already exist: lightweight machine learning models running locally on community devices; mesh networks that distribute computation without centralized servers; architectures where local communities develop situated AIs based on small, specific datasets. These systems can federate while maintaining autonomy, creating cooperative networks rather than monopolies. The goal is not a "green AI" within capitalism, but a break in the link between artificial intelligence and extractive accumulation.

Incorporating subaltern knowledge. The datasets training dominant AI reflect colonial, patriarchal, and racist epistemologies. Yet decolonial projects are already underway: from digital archiving initiatives managed by Afrodescendant communities rejecting Western museological categories, to feminist tech collectives building facial recognition algorithms trained on non-white, non-binary bodies to subvert racial and gender bias embedded in commercial systems. Transforming "errors" and deviations from the dominant norm into tools of resistance against standardization turns technology into an epistemological battlefield, not merely a technical one. A striking example is Masakhane⁵⁰, a pan-African network developing AI models for marginalized African languages, adapting the machine to local knowledge rather than the other way around.

These practices are fragmented, precarious, often marginal, yet they are situated technologies, responding to shared needs, capable of amplifying community practices without reproducing extractive logics. The goal certainly cannot be to "humanize" AI or make it "ethical" within the existing system, but to wrest infrastructures from oligarchic control and transform them into tools of struggle, embedded within the uprisings already spreading across the planet. Our task is not to reform the megamachines of capital — even if find ourselves forced, here and there, to tactically reclaim them — but to sabotage their cognitive mechanisms and build, in their interstices, new architectures.

As in quantum entanglement, where distant states influence each other instantaneously, a form of connectivity emerges that escapes the deterministic logic of centralized control. Artificial intelligence, in this sense, is one of the battlegrounds on which the possibility of redirecting historical time itself is at stake. 

Before it is too late.⁵¹

Translated from the French by Max Mansoubi

Images: Thaddé Comar