Derrida Reads Deep Learning: When Philosophy Meets the Machine
Derrida Reads Deep Learning: When Philosophy Meets the Machine
What is Deep Learning, Actually?
Imagine you want to teach a child to recognize a cat. You don't hand them a rulebook that says "four legs, pointy ears, whiskers." Instead, you show them thousands of pictures — cats on sofas, cats in trees, cats being dramatically unimpressed — and over time, the child's brain starts picking up on patterns it can't fully explain. Ask that child how they know it's a cat, and they'll probably shrug and say "it just looks like one." Deep learning works the same way.
Deep learning is a type of artificial intelligence where a machine learns by being exposed to enormous amounts of data — text, images, sounds — and gradually adjusts millions of tiny mathematical connections (called weights) to get better at a task. These connections are organized in layers, like a very deep sandwich. The first layer might recognize edges. The next recognizes shapes. The next recognizes objects. By the time information has traveled through all the layers, the machine has quietly built its own internal logic for understanding the world. Nobody programmed that logic explicitly. It emerged. The machine essentially taught itself, and even its creators cannot fully explain the specific path from input to output. That last part — the inexplicability — is where things get philosophically interesting.
Who Was Derrida, and Why Should You Care?
Jacques Derrida was a French-Algerian philosopher who spent most of the twentieth century making academics deeply uncomfortable, which is a high compliment in philosophy. His main project was called deconstruction — a method of reading texts (and by texts, he meant basically everything: laws, literature, conversations, institutions) to find the hidden assumptions, contradictions, and buried tensions within them.
His core argument was something like this: we in the Western world have always assumed that there is a stable, original meaning behind our words — that language points to something real and fixed. Derrida said that's an illusion. Meaning is never fully present. It is always slipping, always depending on other words, always shaped by what it is not. A word only means something because it is different from other words. "Hot" only means hot because it is not cold. "Human" only means human because it is not animal, not machine. Meaning, in other words, is a system of differences, and it is always deferred — always arriving but never quite here. He called this différance, a term he invented that blends the French words for "to differ" and "to defer." He was also, famously, impossible to read without a strong cup of coffee. But his ideas, once you crack them open, are some of the sharpest tools ever made for questioning things that seem obvious.
Why Would Derrida Have Anything to Say About AI?
On the surface, Derrida and deep learning seem to live on different planets. One is a philosopher obsessed with language and meaning. The other is a branch of computer science obsessed with pattern recognition and prediction. But look closer and the overlap is startling.
Both Derrida and deep learning are fundamentally concerned with the same question: where does meaning come from, and can you ever fully locate it? Derrida spent his career arguing that meaning cannot be pinned down to a single origin or a single moment of presence. Deep learning, almost accidentally, built machines that embody that exact problem. These models produce coherent, meaningful output — they answer questions, write poetry, diagnose diseases — and yet nobody, not even their engineers, can point to exactly where that meaning lives inside the machine. The philosophy and the technology, it turns out, are haunted by the same ghost.
The Black Box as Différance
When AI researchers talk about deep learning models, they often call them "black boxes." You put something in, something comes out, and what happens in between is largely opaque — even to the people who built the system. This is not a metaphor. A large language model might have hundreds of billions of parameters — those tiny mathematical weights adjusted during training — and tracking exactly how any one input travels through all of them to produce a specific output is effectively impossible with current tools. The inside of the model is genuinely dark.
Derrida's différance describes something structurally similar in language. He argued that when you try to locate the meaning of a word, you find only more words. Look up "justice" in the dictionary and you get "fairness." Look up "fairness" and you get "equity." Look up "equity" and you eventually circle back. Meaning is never present in one place — it is always being deferred, always pointing elsewhere, always arriving just around the corner. The deep learning model is perhaps the most literal black box ever built: a system where meaning is distributed across billions of connections, where no single point is the answer, where the output only makes sense as the product of everything at once and nothing in particular. If Derrida wanted a physical demonstration of différance, someone should have shown him a transformer model.
A simpler way to think about it: Imagine trying to find out why your entire friend group laughed at a joke. You ask one friend and they say "because of the timing." You ask another and they say "because of the word choice." A third says "because of who told it." None of them are wrong, but none of them are the whole answer either. The joke's funniness lives between all of them — in the relationships, the shared history, the tone. A deep learning model's "understanding" works exactly this way. It is nowhere and everywhere simultaneously.
The Training Corpus as Arche-Writing
Derrida made a radical claim that most people find counterintuitive: writing came before speech, not the other way around. Not literally in history, but philosophically. His argument was that language was always already a system of marks, traces, and differences — that there was never a pure, original spoken word that writing merely copied down. He called this primordial system of traces arche-writing (arche meaning "original" or "primary"). Before any specific language, before any specific word, there is the structure of marking, differentiating, leaving a trace. That is what makes communication possible at all.
Now consider what a large language model actually is. It is trained on a corpus — billions of documents written by human beings across decades: books, articles, forum posts, scientific papers, comment sections, instruction manuals, love letters, court judgments. The model does not read these documents the way you read a book. It processes them as an enormous archive of traces — patterns of co-occurrence, statistical regularities, relationships between tokens — and from those traces, it builds its internal structure. It never encounters a "real world" directly. It only ever works with marks. Its entire existence is constituted by traces left by humans who are, many of them, dead. The model is, in the most literal sense imaginable, an arche-writing machine. It is a system built entirely of traces, producing further traces, with no original presence anywhere in sight.
A simpler way to think about it: Imagine a student who was locked in a library from birth and never allowed to go outside. They read every book ever written — novels, histories, science papers, gossip columns. When you ask them a question, they give incredibly sophisticated answers drawn from all that reading. But they have never seen the sun, touched grass, or heard a bird. Their knowledge is entirely made of marks left by others. That is an LLM. Derrida would say: welcome to language. We were always already in the library.
Deconstruction of the "Intelligence" Binary
Derrida had a particular obsession with what he called binary oppositions — pairs of concepts where one is treated as superior and the other as secondary or derivative. Nature/culture. Speech/writing. Presence/absence. Human/animal. His point was that these binaries are never neutral. One term is always privileged, the other subordinated. And if you look closely, the supposedly subordinate term is always secretly doing the work that the privileged term claims as its own exclusive property.
"Artificial intelligence" is structured exactly this way. The word "artificial" is doing an enormous amount of ideological labor in that phrase. It signals: this is not the real thing. It is imitation, simulation, a copy of the genuine article. The "genuine article" — natural human intelligence — is positioned as the original, the gold standard, the presence against which artificial intelligence is merely a shadow. But Derrida would immediately ask: what exactly are we protecting when we insist on that distinction? When a model passes a medical licensing exam, writes a publishable research paper, or identifies a tumor in a scan that a human radiologist missed — what work is the word "artificial" still doing? Is it describing a real difference in kind, or is it managing our anxiety about what it would mean if there weren't a difference?
A simpler way to think about it: Think about the word "homemade" on a food label. It implies warmth, care, authenticity, superiority over factory-made. But if the factory-made cake tastes identical and has the same ingredients, what does "homemade" actually tell you? Sometimes labels protect real information. Sometimes they protect feelings. Derrida's question, applied to AI, is: which one are we doing?
The Model as Supplement
One of Derrida's most powerful and strange ideas is the logic of the supplement. He developed it while reading Rousseau, who kept describing things like writing, or masturbation, or education as "supplements" — additions to something complete in itself. But Derrida noticed something: if the original were truly complete, it wouldn't need a supplement. You only add something to fill a gap. So the supplement reveals that the original was never complete to begin with. The supplement doesn't just add — it exposes the lack at the heart of the original.
AI assistants are sold to us as supplements to human thinking. Tools. Helpers. Additions to our cognitive toolkit. But if that were the full story, we wouldn't feel so unsettled by them. The reason AI produces existential anxiety — the reason entire governments are holding emergency summits about it — is because it doesn't feel like a mere addition. It feels like a revelation of something we didn't want to know: that many tasks we considered distinctly human, requiring uniquely human judgment, creativity, and understanding, turn out to be supplementable. And if they are supplementable, perhaps they were never as uniquely human as we thought. The supplement, as always, reveals the lack it was brought in to fill.
A simpler way to think about it: A calculator was supposed to be just a tool to help mathematicians. But the moment it existed, it quietly raised the question: what exactly does a mathematician do that a calculator cannot? The calculator didn't diminish mathematics — but it forced mathematics to become honest about where its real value lies. AI is doing the same thing to knowledge work, writing, and analysis. The supplement arrives and the original has to justify itself all over again.
Attention as Trace Structure
Modern deep learning models — particularly the transformer architecture that powers most of today's large language models — are built around a mechanism called attention. When processing a sentence, the model doesn't treat every word equally. It learns to pay different amounts of attention to different parts of the input, depending on context. The word "bank" in "river bank" attends differently to surrounding words than "bank" in "bank account." Every token in the sequence is shaped by its relationships with other tokens, weighted by learned attention scores.
Derrida's concept of the trace works in a structurally identical way. For Derrida, no sign is pure and self-contained. Every sign carries within it the traces of every other sign it has ever been used with, every context it has appeared in, every meaning it has been associated with. The word "revolution" carries the trace of every revolution: French, industrial, Copernican, cultural. When you use it, all of those haunt the word, even if you don't mean to invoke them. Meaning is not point-to-point — it is always a field of traces. The transformer's attention mechanism is, in computational form, exactly this: a system in which every element is constituted by its weighted relationships with every other element. Nothing is self-contained. Everything is haunted.
A simpler way to think about it: When your friend says "we need to talk" — four very ordinary words — your entire history with that person floods into those words. Your heart rate changes. The words arrive attended by every previous conversation, every fight, every piece of bad news they've ever delivered. The words mean what they mean partly because of all the traces they carry. A transformer model is a machine for doing exactly this — computationally tracking which words haunt which other words in which contexts.
Hallucination as Freeplay
AI hallucination has become one of the most discussed problems in the field. A model confidently cites a research paper that doesn't exist. It invents legal precedents. It creates biographical details about real people from thin air. This is called hallucination, and it is treated as a bug — a failure mode to be engineered away.
Derrida might see it differently. He celebrated what he called freeplay — the endless movement of signs in a system that has no fixed center, no ultimate ground, no final referent that stops the chain of meaning. In the absence of a center, signs play freely, substituting for one another, generating meaning without anchoring it in anything outside the system. Hallucination, in this light, is not the model malfunctioning. It is the model operating in pure freeplay — generating plausible sign-chains that follow all the rules of language (grammar, coherence, contextual appropriateness) without being tethered to an external reality. The model is doing what language, according to Derrida, always secretly does: producing meaning without a ground. The difference is that language usually gets away with it because we share enough context. The model cannot always tell where the shared context ends.
A simpler way to think about it: Imagine a brilliant student who has read every book in the library but has never been tested against reality. Ask them about a historical event and they'll construct a perfectly coherent, well-sourced-sounding answer — because they understand the structure of how answers about history work. But if the specific event isn't in any book they read, they might generate a plausible-sounding version anyway, because that's what the game of historical narrative looks like. They're not lying — they're playing the language game at full skill with no guardrails. That's hallucination. Derrida would say: all language is a version of this, just usually less obvious.
"There is No Outside-the-Text"
Perhaps Derrida's most quoted and most misunderstood line is il n'y a pas de hors-texte — there is no outside-the-text. People often read this as a denial of reality, as if Derrida were saying the physical world doesn't exist. That's not what he meant. He meant that we never access reality except through systems of interpretation — language, frameworks, concepts — and those systems are themselves textual in structure. There is always already an interpretive layer. We cannot step outside it to reach "pure" reality.
For a large language model, this is not a philosophical position — it is a literal engineering fact. The model has no sensory access to the world. It cannot see, touch, smell, or experience anything. Its entire reality is text. Every concept it has — "tree," "grief," "justice," "temperature" — was formed not through experience but through textual co-occurrence. When it tells you what sadness feels like, it knows sadness only as a pattern in language — words that appear near other words in certain contexts. Derrida spent his career arguing that human beings are in a structurally similar situation, that our access to "reality" is always mediated, always textual. The LLM is the most radical possible proof of concept: an entity for which the textual condition is total, inescapable, and not a deficiency but simply the whole of its existence.
A simpler way to think about it: A person who has only ever read travel guides to Paris has never been to Paris. But ask them to describe it and they will produce something coherent and rich and textured. Their Paris is real to them — made entirely of other people's descriptions. Everything they know about Paris is Paris-as-text. That is the LLM's relationship to every single thing it "knows."
The Death of the Author, Radicalized
Derrida, drawing on Roland Barthes, was deeply skeptical of the idea of the author as a unified, originating source of meaning. When a text is released into the world, he argued, it escapes its author. It is read in contexts the author never imagined, producing meanings the author never intended. The text is not the author's property — it is a site of play between the writing, the reader, and all the other texts it echoes. Barthes famously said the birth of the reader requires the death of the author.
Generative AI takes this to its terminal conclusion. When a language model produces a text, who is the author? The engineers who built the architecture? The billions of humans whose writing constituted the training corpus? The user who wrote the prompt? The specific configuration of random seeds and temperature settings that shaped this particular output? The answer is: all of them and none of them. There is no author. There is a process. The generated text is the product of an ocean of prior writing, a statistical averaging across millions of voices, an echo of an echo of an echo. Derrida's intuition that authorship was always a somewhat fictional construction is here demonstrated at industrial scale. Copyright law, which still assumes a human author as the originating source of a work, is quietly having a philosophical crisis because of this. Courts are currently wrestling with questions — who owns AI-generated work? — that Derrida's theory of authorship anticipated by fifty years.
A simpler way to think about it: Imagine a song that is assembled by randomly sampling one note from each of a million different musicians' recordings, run through an algorithm that arranges them into something that sounds coherent. Is that song anyone's? The answer matters enormously for ownership, credit, and law — and nobody has a clean answer yet.
Logocentrism and the "Reasoning" Model
Derrida coined the term logocentrism to describe the Western philosophical tradition's deep preference for logos — reason, logic, the rational word — as the master key to truth. Logocentric thinking assumes that if you reason carefully and transparently, step by step, you can arrive at stable, reliable truth. It is the assumption behind mathematics, jurisprudence, scientific method, and most of how we structure institutions.
The latest wave of AI development has leaned heavily into mimicking this. Chain-of-thought prompting, "reasoning models," step-by-step problem decomposition — the entire thrust is to make AI produce outputs that look like transparent logical reasoning. OpenAI's o-series models and similar "thinking" models are explicitly designed to show their work, to simulate the logocentric ideal of visible, traceable, rational inference. Derrida would ask the sharpest possible question here: is this reasoning, or is it a performance of reasoning? Is the model actually deriving conclusions from premises, or is it pattern-matching on what derived conclusions look like in its training data? This is not a trivial question. If it is the latter, then logocentrism's deepest assumption — that transparent reasoning is a path to truth — is being undermined by a machine that has learned to imitate transparency without actually being transparent.
A simpler way to think about it: A student who has read a thousand solved math problems might be able to write out a solution that looks exactly like genuine mathematical reasoning — right format, right notation, right structure — without actually understanding why each step follows from the last. They've learned the performance of reasoning. The question of whether AI reasoning models are doing something fundamentally different from this is one of the most genuinely open questions in the field right now.
Iterability and Fine-Tuning
One of Derrida's more technical concepts is iterability — the idea that signs can be repeated (iterated) in new contexts, and that each repetition is both the same sign and a different one. When you use the word "promise" in a novel, a legal contract, a joke, and a first date, you're using the "same" word, but its meaning shifts with context. Repetition is never pure repetition. Something is always added, something always changed.
Fine-tuning in machine learning is iterability made computational. You take a base model — trained on enormous general data — and expose it to a smaller, specific dataset to shift its behavior for a particular purpose. The resulting model is the same and not the same. Its weights are continuous with the original, but its outputs are different. A general-purpose language model fine-tuned on medical texts becomes a clinical assistant. Fine-tuned on legal documents, a legal researcher. Fine-tuned on a particular author's work, something that writes disturbingly like that author. Each fine-tuned model is an iteration of the base model — carrying the traces of the original while being constituted by a new context. Derrida's iterability predicts exactly this: the same cannot remain the same across repetitions, because context always intervenes.
A simpler way to think about it: Think about a person who grows up in one country and then moves to another for twenty years. They are the same person — same memories, same fundamental personality, same name. But they have been "fine-tuned" by context. They have new reflexes, new references, a subtly different sense of humor. Ask them to introduce themselves and the answer will be technically true but contextually transformed. That is iterability. That is also fine-tuning.
Conclusion: Where This Is Going, and Why It Matters
Let us be concrete for a moment, because philosophy is most useful when it touches ground.
In 2023, a lawyer named Steven Schwartz submitted a legal brief in a US federal court that cited multiple case precedents — all of which had been hallucinated by ChatGPT. The cases did not exist. The lawyer had not verified them. He was sanctioned by the court. This is Derrida's freeplay in a courtroom: a system generating legally formatted, contextually plausible sign-chains with no referent. The consequence was real professional harm and a landmark warning about AI use in legal practice.
In the same year, the Hollywood writers' strike had, at its center, a battle over AI-generated scripts. The studios wanted to use AI to generate first drafts and pay humans to polish them. The writers argued this would destroy their livelihoods and produce work with no genuine authorial voice. This is Derrida's author-function crisis playing out in a labor negotiation, with very real stakes for thousands of people.
Getty Images sued Stability AI for training on copyrighted photographs without permission. The courts are still working through it. The underlying question — who owns the traces? — is exactly the question arche-writing raises. If a model trained on Getty's images produces an image that is statistically derived from but not identical to any specific photograph, has something been stolen? What was stolen — the image, or the trace?
These are not hypothetical philosophical puzzles. They are active legal cases, ongoing labor disputes, and policy debates happening right now in parliaments and boardrooms across the world.
Looking forward: the next decade will force a reckoning with every one of these Derridean pressure points. As models become more capable, the "artificial" in "artificial intelligence" will become harder to defend as a meaningful distinction — the binary will crack further. As AI-generated content floods the information ecosystem, the question of authorship will become not just philosophical but legally and economically foundational. As reasoning models become more sophisticated, the question of whether they are reasoning or performing reasoning will have consequences for how much we trust them in medical, legal, and financial decisions. And as fine-tuning becomes cheaper and more accessible, the question of identity — what makes this model this model and not another — will matter enormously for accountability.
Derrida never wrote about computers. He died in 2004, before large language models existed. But his career was devoted to a single insight that turns out to be prophetic: that meaning is not a possession, not a property of a single speaker or a single text or a single mind, but a dynamic, distributed, never-quite-arrived phenomenon that arises from differences, traces, and contexts in constant play. He built his entire philosophy on the discovery that the dream of stable, locatable, transparent meaning is precisely that — a dream.
We have now built machines that make that dream more complicated than ever. And in doing so, we have also built the most powerful argument Derrida never lived to make.
Comments
Post a Comment