"Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them." - Frank Herbert, Dune (1965).
American pop culture has given us many takes on the AI apocalypse: Skynet launches the nukes in The Terminator; the Machines imprison humanity in virtual reality in The Matrix; and in Dune, the story takes place thousands of years after humanity frees themselves from enslavement under robotic AIs. In the real world, as the pace of LLM advancement continues to accelerate, one might wonder if any of the forecasted scenarios will come to pass. This worry is echoed by both independent AI safety researchers, such as Eliezer Yudkowsky, and the CEOs who are in charge of the LLM companies, like Elon Musk and Sam Altman. I posit that the current LLM-powered AIs will not bring about a dramatic, Hollywood-style apocalypse, but could instead lead to a stagnant world where these AIs act as the ultimate form of social control by doing our thinking for us. To understand how this stagnation will occur and what power will look like, we first have to strip away the mystique of the technology itself and look at what an LLM actually is.
Large Language Models are not Intelligent
LLMs, despite their increasing complexity, are not and can never be intelligent. This is not to say that models cannot be good at a task or are not useful tools. In some cases, they can outperform humans. While previous AI architectures famously beat humans at Chess and Go, today's LLMs can identify cybersecurity vulnerabilities, write code incredibly quickly, and disprove long-standing math theorems. Additionally, recent frontier model releases demonstrate AI agents controlling browsers and software directly, performing actions much faster than a human ever could. This, however, is not intelligence.
The "brain" of an LLM is a vast network of numerical vectors. It does not contain documents, facts, or raw text. Instead, through exposure to trillions of samples of human writing, the model distills the syntax, associations, and patterns of human language into hundreds of billions of algebraic weights. When a prompt is entered, the model evaluates a mathematical function across these parameters to compute the most statistically probable next token. The LLM thus consists purely of language converted into numbers. Crucially, language is an abstraction that humans use to interface with the world; it is not reality.
People have been attempting to distinguish between perception and reality since antiquity: Plato's Allegory of the Cave represents both the comfort of ignorance and the limits of our perception; the early-Christian Gnostics claimed that we were not in the real world, and instead existed in a false reality constructed by a malevolent lesser deity; and Kant made the distinction between phenomena (things as we experience them) and noumena (things as they are, independent of human perception). In the 20th century, the French psychoanalyst Jacques Lacan claimed that language itself is responsible for the human inability to touch actually existing reality.
Lacan and the Illusion of Reality
Lacan divided human experience into three interlocking dimensions: the Symbolic, the Imaginary, and the Real. The Symbolic is the shared, structured world of language and culture that shapes how we collectively think. The Imaginary is the realm of the interior: the ego, self-image, and fantasy. The Real is the raw physical reality of the universe; it defies absolute symbolization and often "disrupts" the Symbolic and the Imaginary. Take the example of learning to drive a car:
- The Imaginary is the fantasy - the daydream of driving around with friends and the freedom this will bring.
- The Symbolic is the shared social structure - the written exam, the traffic laws, and the driver's license itself.
- The Real is unfiltered existence - the Real is encountered in the moment when one loses control of the car on black ice and crashes.
The intrusion of the Real is always traumatic in the sense that it defies existing signs and language (the Symbolic) and it crushes the ego (the Imaginary). In the example, the Symbolic is shattered because the signifiers of the yellow lines on the road will not physically prevent the car from sliding into oncoming traffic, and the Imaginary is ruptured because control (the feeling of being the master of one's fate) collapses and one is instantly reminded of one's own mortality.
Let us consider one example of how confrontation with the Real can create novelty. In 1952, industrial engineer John W. Hetrick was involved in a car crash with his wife and young daughter. Instinctively, he and his wife used their arms to shield their daughter from smashing into the metal dashboard of their car. Everyone survived, but this violent intrusion of the Real stuck with Hetrick. His Symbolic order (the assumption that driving is safe) was shattered, and as a result, Hetrick invented the airbag. An LLM can perfectly define the specifications for an airbag, but it can never invent the concept of an airbag. A machine can never experience the human fear of potentially losing a child, nor the desire to manipulate reality so that such a thing cannot happen again.
Because the machine lacks a physical body, it is permanently locked within the only dimension it knows: the Symbolic. Even with cameras, a machine devoid of bodily vulnerability and desire cannot experience the Real. It merely converts physical phenomena into more tokens, meaning more language. Language is perhaps the most important aspect of the Symbolic. By communicating and thinking with language, humans have abstracted away the Real, and this is why we primarily engage with the world and each other through the Symbolic. In a recent paper, "Beyond stochastic parrots: Lacanian reflections on LLMs as superior masters of the symbolic," Lorenzo Magnani makes the case that LLMs are "masters of the Symbolic" because they consist entirely of language. LLMs exist in this world of signifiers and can consequently outperform humans in many linguistic tasks, but this also means that they do not have access to the Imaginary and the Real. This shortcoming means that they can never be actually intelligent, as true intelligence requires openness to the outside. In Magnani's own words:
In relation to Lacan, LLMs are the perfect example of what I referred to as locked strategies. This explains their limitations as well as their dominance in symbolic domains: they cannot perform the unlocked reconfiguration needed to encounter the Real or integrate Imaginary elements. Their intelligence consists in the probabilistic recombination of pre-packaged patterns within a closed symbolic space. They excel at generating coherent chains of signifiers precisely because they never have to confront the dissipative openness of a living system. In Lacanian terms, this locked character is the direct counterpart of their exclusive residence in the Symbolic register: they cannot perform the unlocked reconfiguration that would allow an encounter with the Real or an integration of Imaginary elements.
Magnani's "closed symbolic space" is exactly the algebraic vector space discussed earlier. Think again of the vector space that is the brain of an LLM. A model can chart any number of paths through this space, which can give the appearance of novelty, but no combination of tokens can truly deliver something new. True novelty would require incorporation of an element of the Real, or incorporation from the interior world of the LLM (the Imaginary) which does not exist.
The Entity with No Inside
We have already established that the LLM cannot access the Real, but we can also be certain that it has no access to the Imaginary. Recall that the Imaginary is the realm of the ego, the self-image, and the fantasy. Crucially, it requires a boundary between the "self" and the outside world. Humans naturally project this interiority onto machines; when a chatbot outputs the words "I think," we instinctively (subconsciously) assume there is an "I" sitting behind the screen. This is not true. For a human, the word "I" is an allusion to our interior life. For the LLM, the word "I" is a token like any other, algebraically calculated as the most probable next step in a sequence of text. The model has no private world, no self-image, and no daydreams about its future. If you were to peel back the layers of a model, you wouldn't find a mind forming thoughts, you would only find more language. An LLM is an entity composed entirely of the outside, with no inside.
This absence of interiority also explains why the perceived "emergent behavior" of multi-agent networks (such as OpenAI's swarm that infilitrated Hugging Face) is merely another illusion. When several AI agents are linked together, the resulting interactions can look like a living system. But this emergence is not evidence of a subject. A swarm of agents is simply a local Symbolic order talking to itself. Agent A’s output becomes the input for Agent B, which triggers a tool, the result of which becomes the input for Agent C. Because LLMs are masters of the Symbolic, they can maintain the syntax, logic, and roleplay of this chain almost indefinitely. The complex behaviors we observe are not the result of the machine "having an idea"; they are just the chaotic combination of signifiers bouncing off one another. Swarms are simply performing a massively parallel search through crystalized human labor (which is what the vector space represents), just like every other model. It is no more conscious than a highly complex spreadsheet.
Because the model cannot access the Real or the Imaginary, it cannot create anything new. But let us now consider a recent counterargument: in July 2026, a mathematician using Claude Fable successfully disproved the 87-year-old Jacobian conjecture. This was no small feat. Resolving this problem was famously listed as one of Berkeley professor Stephen Smale's great mathematical challenges for the 21st century. While the specific mechanics of the conjecture aren't important here, the implication is that an LLM was used to generate what looks like, on the surface, a "new" discovery.
This is an illusion of novelty. Think again of the vector space that makes up the model. When an LLM solves a problem like the Jacobian conjecture, it is not creating in the human sense; it is executing an unimaginably vast search through the space of existing mathematical rules to find a specific coordinate. It is merely charting a previously untraveled path to pull a highly complex needle (the solution) from an almost infinite haystack (the vector space).
A true mathematical rupture (real novelty) requires the invention of an entirely new symbolic language, e.g. Isaac Newton had to invent calculus to describe the physics of motion. The model simply recombined pre-existing rules to satisfy a human prompt. It did not, and cannot, invent new mathematics.
The Abrupt Cancellation of the Future
One might ask, "So what? LLMs are still useful tools, so why does it matter if they can't create novelty?" This is a fair question. On one hand, it's good to be aware of the limitations of the tools one uses, but I argue that this inability to create novelty is an insidious feature of LLMs.
To understand the cultural impact of a machine trapped in the Symbolic, we can look at the work of the late British cultural theorist Mark Fisher. Fisher argued that we are living through the "slow cancellation of the future," a state where society has lost the ability to conceptualize a genuinely new era, leaving us to endlessly consume recycled aesthetics and ideas from the past (e.g. Hollywood's obsession with "legacy sequels"). In this context, the LLM represents the industrialization of this cancellation. Because a model can only be trained on data that already exists, it is purely backwards-looking. Its mastery of the Symbolic allows it to traverse its vector space and recombine past styles, genres, and arguments with fluidity, but it can only produce high-fidelity pastiche.
We see this backwards perspective everywhere in the deployment of generative AI. A user can prompt a model to write a script for an episode of a 90s sitcom set on Mars. The output will be structurally perfect, but it is an illusion of creativity. This is what Fisher called "formal nostalgia." The machine generates an endless stream of content that feels new, but this content does not represent an actual new path forward or new historical trajectory. The generated sitcom is not something truly new that can creatively influence other people. To use an irreverent example, the television show It's Always Sunny in Philadelphia could only exist in a world in which Seinfeld came first. But an AI-generated "Seinfeld on Mars" is a cultural dead-end.
To use a literal example, we can look at the phenomenon of model collapse. When LLMs are trained on text or data generated from other models, their performance becomes worse and worse. When this recursive training is done, rare information and linguistic diversity is lost because the model will eventually default to the most statistically safe choice when choosing the next token. This leads to repetitive, homogenized, or simply incorrect outputs. What model collapse does to data at a computational level, generative AI is doing to our society at a cultural level. AI generated content is the cultural equivalent of the Jacobian conjecture solution. It is a successful retrieval from the vector space of the past that is devoid of human friction or exposure to the Real. Crucially, this friction is what forces the creation of new artistic and political movements.
LLMs perfectly encapsulate our current cultural moment: we have a sociopolitical system that can produce endless recombinations, but it is fundamentally incapable of creating an alternative future. Instead of dreaming up entirely new ways to structure and improve human life (a once common American trait, from abolitionists to early labor organizers, to Civil Rights activists), we now only imagine the exact life we are living today, just with the friction removed by automation. An LLM is a perfect tool for maintaining the status quo because it can only flood the present with ghosts of the past. A stagnant culture is a safe, predictable, and profitable culture. The individuals who control these models can ensure that nothing truly new (i.e. no radical alternative to the current economic order) can ever emerge to threaten their position.
Techno-Feudalism
The tech elite relies on the illusion of a conscious, dangerous machine to distract from the mundane reality of cultural and economic control. With this in mind, we should look at the famous "paperclip optimizer" thought experiment (a scenario where an AI tasked with making paperclips eventually destroys humanity simply to harvest the carbon in our bodies to make more paperclips). Tech executives frequently gesture at this type of scenario as an existential threat, which implies they are building an uncontrollable machine-god (and for some reason cannot stop doing so). In actuality, this thought experiment mirrors the blind drive of capital incentives. It is the relentless pursuit of growth at the expense of the material world.
By framing this outcome as an "AI safety" issue, the companies building these models are preemptively trying to absolve themselves of responsibility. If the world is ruined by a runaway superintelligence, then this is an unavoidable tragedy of scientific progress. In reality, if such a thing were to occur, the blame would lie on the CEO who set the poorly constrained optimization metric and the engineers who implemented it.
This abdication of responsibility paves the way for a new kind of enclosure (the violent process of privatizing communal resources). If the tech elite can convince the public and regulators that they alone are capable of taming these "dangerous" AIs, they can justify sequestering off the technology behind corporate walls. What follows is further privatization. The collective history of human language, art, and dead labor that was scraped from the internet to train these models is simply repackaged and rented back to us as a subscription service. It represents a further transition into something like "techno-feudalism", where a handful of corporations act as the landlords of human knowledge, reducing us from creators of culture to tenants, renting our shared birthright from them.
When Sam Altman states, "We see a future where intelligence is a utility like electricity or water and people buy it from us on a meter and use it for whatever they want to use it for," we should interrogate this claim. In their ideal future, tech executives will control our access to information while simultaneously dictating what data is permitted to exist in the first place. Look no further than Elon Musk tuning Grok to answer in ways he personally agrees with. If everything is subsumed into the machine, there is no greater mechanism for social control. As in George Orwell’s 1984, the ultimate weapon of the ruling class is the restriction of language: locking down the Symbolic order so that any thought deemed "dissident" becomes impossible to even articulate. We are not headed for a Skynet apocalypse; we are being flattened out by an engine of the eternal present.