Spoiler alert: Plato did not write about large language models 2,400 years ago.
But he did imagine a group of people chained inside a cave, staring at images projected onto a wall and mistaking those images for reality. Today, billions of us spend hours staring into illuminated rectangles filled with people we have never met, places we have never visited, and events we cannot verify.
At least Plato’s shadows were cast by real objects.
Ours can be generated by a GPU.
The travel vlogger walking through Peru may never have left their bedroom. The model selling subscriptions on Instagram may not exist. The politician’s voice may have been cloned. Most of the time, we only notice when the illusion breaks: six fingers, mangled text, a reflection that follows the wrong laws of physics.
The glitch is the moment the prisoner sees the wall.
And the comparison gets stranger. Generative AI does not merely produce shadows. Inside its models, concepts are organized as mathematical relationships in high-dimensional spaces. Chair. King. Woman. Justice. Beauty. Not as physical objects or lived experiences, but as coordinates inferred from patterns in human data.
Plato called his abstract structure the Forms.
AI researchers call theirs latent space.
They are not the same thing. But they are close enough to create a deeply uncomfortable question: did we use computation to discover a mathematical structure beneath reality, or did we build the most convincing cave in human history?
The Original Cave
Plato presents the Allegory of the Cave in Book VII of The Republic.
Imagine prisoners chained inside a cavern from childhood. They cannot turn their heads. Behind them burns a fire. Between the fire and the prisoners, puppeteers carry statues of people, animals, and objects along a raised walkway.
The fire casts shadows onto the wall.
The prisoners give those shadows names. Dog. Tree. Person. Because the wall is all they have ever seen, they do not understand the shadows as representations of something else. The image is the object. The echo is the voice. The projection is reality.
Then one prisoner is freed.
He turns toward the fire and the objects casting the shadows. The light hurts. What he had called reality suddenly looks thin and fraudulent. When he is dragged outside, the sun blinds him again. Slowly, painfully, he learns to see the world that produced the images.
For Plato, this is a story about education, knowledge, and the difference between appearance and truth. The sensible world gives us partial and unstable impressions. Reason points beyond those impressions toward what Plato called the Forms: the intelligible, unchanging structures that make knowledge possible.
In other words, the chair in your kitchen can break, rot, or lose a leg. The concept of a chair does not.
No two chairs are identical. A plastic patio chair, a Victorian throne, and a 1970s beanbag barely resemble one another. Yet we place them inside the same category without much effort. Plato’s answer was that particular objects participate, imperfectly, in a stable Form.
The physical chair changes.
Chair-ness remains.
Plato believed knowledge meant turning away from the shifting shadows and toward the structure behind them. Generative AI appears to do something similar. It consumes millions of particular examples and compresses their relationships into mathematics.
But appearances can be deceiving.
Especially when appearances are the only thing the machine has ever known.
The Forms Inside the Machine
An AI model has never sat in a chair.
It has no sore back. No memory of cheap classroom plastic. No instinctive understanding that a chair is something a tired body collapses into after work.
It has data.
Images. Captions. Text. Pixels. Tokens. Patterns left behind by people who have encountered chairs in the world.
During training, a model learns statistical relationships within that data. Similar features and concepts become organized near one another in a high-dimensional representational space. Dissimilar things move farther apart. The model does not store one perfect JPEG labelled CHAIR. It learns enough of the recurring structure to generate a new chair that never physically existed.
This is where the Platonic comparison becomes seductive.
The model looks across countless imperfect examples. It ignores many of their accidental details. The paint. The dust. The room. The angle of the camera. What remains is a mathematical structure capable of producing another recognizable example.
A human sees a chair and thinks of sitting.
The machine sees a position relative to other positions.
One of the most famous demonstrations comes from word embeddings. In these systems, words are represented as vectors: long lists of numbers placing them inside a mathematical space. Researchers found that some relationships could be approximated through arithmetic:
king − man + woman ≈ queen
The result is not magic, and it does not prove that a machine has discovered the eternal essence of royalty. It shows that the training data contains a relational structure the model can encode geometrically. Gender, status, geography, tense, and countless other patterns become directions through a space.
That distinction matters.
Plato’s Forms are supposed to be more real than the objects we perceive. An embedding is downstream from human language. It inherits our categories, associations, omissions, and prejudices. Plato’s Form of Justice is an objective standard. An AI model’s representation of justice is a statistical fossil made from everything people have written about it.
One aims at truth.
The other predicts what comes next.
Still, the resemblance is difficult to ignore. We have taken concepts that once seemed immaterial and rendered their relationships as geometry. Abstract philosophy became something we can probe inside a machine.
We moved the Forms out of heaven and into a data centre.
The Platonic Representation Hypothesis
In 2024, researchers Minyoung Huh, Brian Cheung, Tongzhou Wang, and Phillip Isola gave this comparison a name: the Platonic Representation Hypothesis.
Their claim is not that neural networks have proven Plato right. It is that as different AI models become more capable, their internal representations appear to become more alike. Large vision and language models trained on different data and different tasks can organize relationships between objects in surprisingly similar ways.
The authors hypothesize that these models may be converging toward a shared statistical model of reality.
Think about how strange that is.
One machine learns through images. Another through text. Another through audio. They begin from different sensory shadows, built by different companies with different architectures, yet may form comparable maps underneath.
If every sufficiently capable model starts drawing the same map, there are two possible explanations.
The first is mundane: the models are trained on overlapping human culture, optimized through related methods, and rewarded for capturing the same useful correlations.
The second is Platonic: there really is a common structure beneath the data, and intelligence tends to find it.
The paper presents a hypothesis, not a revelation. Later work has challenged how much of the apparent convergence survives stricter measurement. Some broad similarities may partly reflect model scale, while local relationships remain more convincing.
The map is appearing.
We still do not know whether it belongs to reality or to the machinery we use to measure it.
The Reverse Cave
Plato’s prisoner escapes by moving from images toward their source.
AI moves in the opposite direction.
It begins with records of human experience: photographs, books, conversations, paintings, videos. These are already representations. The model compresses them, then produces new text and images from that compression. When those outputs spread across the internet and enter future training sets, the next model learns from representations of representations.
Shadows trained on shadows.
Researchers have warned that repeatedly training models on generated data can degrade their outputs, a problem commonly described as model collapse. Rare features disappear. Errors compound. The strange edges of reality get replaced by the model’s statistically safe centre.
The average consumes the exception.
This is not only a technical problem. A model trained on the past does not discover Beauty or Justice in Plato’s sense. It learns which faces have historically been labelled beautiful and which decisions have been described as just. Bias does not enter the cave as sabotage. It enters as training data.
No conspiracy is required.
The model reflects the archive. The archive reflects the institutions that preserved it. The institutions reflect the people who had the power to record, publish, classify, and exclude.
Then the output returns to us wearing the authority of mathematics.
We ask the model what a CEO looks like. What a criminal looks like. What a beautiful woman looks like. What a trustworthy voice sounds like. The machine gives us the statistical answer, and the statistical answer produces more images, more stories, more hiring decisions, more expectations.
The shadow becomes evidence for itself.
Plato worried that prisoners would defend the wall because it was the only reality they understood. We may do something worse. We may let the wall update the world until reality begins to resemble its projection.
What If the Shadow Is Perfect?
Right now, AI reveals itself through mistakes.
The sixth finger. The impossible reflection. The confident lie. These glitches reassure us that there is still a meaningful border between the image and the thing it imitates.
But what happens when the errors disappear?
Imagine a generated world where the physics are exact, every voice is convincing, and every interaction responds as a person would. The light hits the water correctly. The stranger remembers your last conversation. Nothing breaks the illusion because the illusion behaves exactly as expected.
Is a perfect shadow still a shadow?
The rationalist answer is no. Gottfried Wilhelm Leibniz dreamed of a universal symbolic language in which disputes could be reduced to calculation. Modern physicist Max Tegmark goes further, arguing that physical reality is itself a mathematical structure. From this view, if a model captured every relationship perfectly, nothing essential would remain outside the map.
The math would not describe reality.
The math would be reality.
David Hume would be less impressed. A machine can learn every statistical relationship between fire, smoke, pain, and heat without ever feeling a flame. It has the pattern without the impression. The description without the sensation.
Immanuel Kant creates an even harder boundary. Human beings never encounter the thing-in-itself directly; we experience the world after it has been organized by our own faculties of perception. AI is trained on the products of those faculties. It is a filter built from the output of another filter.
A goggle for our goggles.
From that view, even a flawless model would remain downstream from reality. It could produce the perfect map of human experience while knowing nothing of what exists beyond it.
So which is it?
Did the machine find the structure of the universe, or the structure of our descriptions of the universe?
The answer depends on whether you believe a perfect representation leaves anything behind.
The Map Over the Territory
Jorge Luis Borges imagined an empire whose cartographers became so precise that they produced a map on the same scale as the empire itself.
One mile of paper for one mile of land.
The map covered the territory it was supposed to represent. Later generations abandoned it, leaving torn fragments across the desert.
Generative AI reverses Borges’s ending. We are not abandoning the map. We are moving onto it.
Our friendships pass through recommendation systems. Our jobs are filtered by scoring models. Our art is scraped into datasets, compressed into vectors, regenerated, and fed back into the culture. Our conversations train systems that will shape later conversations. The representation no longer sits beside the world as a passive copy. It selects what we see and alters what gets made.
The map recommends the territory.
The territory adapts.
The updated territory trains the next map.
Plato wanted the prisoner to turn around, endure the pain of the light, and search for the source of the shadows. We are taking the source material and using it to make the shadows brighter. More responsive. More personal. More comfortable than whatever waits outside the cave.
Maybe AI is approaching a World of Forms: a common mathematical structure toward which different intelligences converge.
Or maybe latent space is only the cave perfected. Not truth, but a model of everything humanity has already mistaken for truth.
Either way, we are no longer just watching shadows on the wall.
We are feeding them.
And eventually, they will feed us back to ourselves.
References
Huh, M., Cheung, B., Wang, T. & Isola, P. (2024). The Platonic Representation Hypothesis. arXiv:2405.07987.
Mikolov, T., Chen, K., Corrado, G. & Dean, J. (2013). Efficient Estimation of Word Representations in Vector Space. arXiv:1301.3781.
Plato. Republic, Book VII, 514a–521b.
Shumailov, I. et al. (2024). AI Models Collapse When Trained on Recursively Generated Data. Nature, 631, 755–759.
Borges, J. L. (1946). On Exactitude in Science.
Gröger, F., Wen, S. & Brbić, M. (2026). Revisiting the Platonic Representation Hypothesis: An Aristotelian View. arXiv:2602.14486.