An LLM isn’t a mirror—it’s more like a hologram, one that reconstructs meaning from language

've long described large language models as mirrors—tools that reflect our language, our questions, and even our cultural biases. It’s an easy metaphor to reach for. When you prompt an LLM, it offers a response that sounds polished and familiar, like a stylized echo of your own thinking. But as these systems grow more powerful, the mirror metaphor begins to collapse. A mirror reflects. It doesn’t interpret, complete, or create.

article continues after advertisement

LLMs do all three.

Ask a model, “What is quantum physics?” and you might get a clean, textbook-like response. “Quantum physics is the branch of science that studies the behavior of matter and energy at the smallest scales, where particles behave both like waves and particles.” It certainly sounds informed. But the model isn’t explaining anything. It’s assembling language based on probability, drawing from patterns in its training data, reconstructing a form that "feels" right. It doesn’t know quantum physics. It simulates the appearance of knowing.

This is why we may need a better metaphor—one that accounts for this generative capacity. LLMs don’t behave like mirrors. They behave more like holograms.

Holograms Don’t Store Images, They Store Possibility

A hologram doesn’t capture a picture. It encodes an interference pattern. Or more simply, it creates a map of how light interacts with an object. When illuminated properly, it reconstructs a three-dimensional image that appears real from multiple angles. Here’s the truly fascinating part: If you break that hologram into pieces, each fragment still contains the whole image, just at a lower resolution. The detail is degraded, but the structural integrity remains.

LLMs function in a curiously similar way. They don’t store knowledge as discrete facts or memories. Instead, they encode relationships—statistical patterns between words, contexts, and meanings—across a high-dimensional vector space. When prompted, they don’t retrieve information. They reconstruct it, generating language that aligns with the expected shape of an answer. Even from vague or incomplete input, they produce responses that feel coherent and often surprisingly complete. The completeness isn’t the result of understanding. It’s the result of well-tuned reconstruction.

article continues after advertisement

Resolution and Reconstruction

What improves as language models scale is their resolution, not necessarily the scope of their memory or the clarity of their output. Larger models have more parameters and finer-grained embeddings. They operate with higher “resolution,” resolving ambiguity with greater precision and representing relationships with more depth. Smaller models can still function, but their outputs are fuzzier, just like a low-resolution hologram. They may miss nuance or falter in maintaining coherence over longer text, but they still generate meaningful responses by drawing from distributed structure.

This explains why even minimal prompts like “Tell me something about love” can yield emotionally resonant replies. The model doesn’t feel love. It reconstructs the shape of language used to talk about love. It draws on patterns found in poems, speeches, essays, and conversations, and it builds an approximation or a "linguistic surface" that feels familiar and complete, even when there’s no lived experience behind it—other than the lived human experience that is "baked" into the training data.

What We Mistake for Intelligence

Our tendency to equate fluency with thought makes LLMs especially persuasive. When language flows with clarity and confidence, we assume there is understanding behind it. Coherence sounds like cognition, but this is a cognitive shortcut that LLMs exploit well. Their outputs are convincing not because they are derived from awareness, but because they so closely mimic the "surface structure" of intelligent speech.

article continues after advertisement

What we’re seeing is a form of epistemological holography (EH). It’s the shaping of knowledge through the structured interference or refraction of language. It's where meaning emerges from patterns of words, not from memory or intent.

EH isn’t deception, it’s construction.

Large language models don’t think or recall. They assemble, drawing on seemingly countless relationships encoded in their training data to generate something that "feels" coherent. And that's because it is coherent, structurally. But coherence isn’t the same as cognition, and fluency isn’t proof of insight. EH helps us better understand what we’re really encountering when a machine appears to “know.”

The Whole in Every Part

This reconstructive ability isn’t just a technical quirk. It may hint at something deeper about intelligence itself, in both machines and minds. In other words, meaning might not be stored in a single place. Instead, it could be distributed, encoded across networks, emerging only when conditions call it forth. Human memory often works this way. When we recall a childhood experience, we’re not retrieving a file—we’re rebuilding a moment from fragments where sensations, impressions, and contextual cues are stitched together to form a coherent whole.

In that sense, our own cognition may be, in part, holographic. We infer, we interpolate, we fill gaps in memory with imagination and expectation. What LLMs do through algorithmic probability, we may do through lived pattern recognition. The comparison is not perfect, but it is provocative. LLMs don’t think like we do, but they may expose something essential about how thought itself is possible.

article continues after advertisement

Seeing the Shape Beneath the Surface

As I've said, LLMs don’t produce knowledge. They produce the form of knowledge—the shape of an answer, the tone of insight, the cadence of a mind at work. But behind that cadence is a statistical mechanism, not a conscious one. They reconstruct meaning the way a hologram reconstructs light—not with substance, but with structure.

This doesn’t mean we should dismiss them. Far from it. These tools are powerful, often beautiful, and increasingly essential. But if we’re going to use them well—if we’re going to integrate them into how we learn, create, and communicate—we need to understand what we’re actually seeing.

Because when we reach out to the hologram thinking it’s solid—when we lean on language and expect substance—we risk falling through, not just in how we engage with machines, but in how we understand ourselves.