Ontology of Information

Here’s a question that sounds simple and isn’t: what is information?

Not data, not knowledge, not wisdom. Information. The thing in the middle. The thing Claude Shannon measured in 1948, the thing Luciano Floridi spent a career trying to define, the thing John Archibald Wheeler thought the entire universe was made of. The thing you’re consuming right now, reading these words.

Everyone uses the word. Nobody agrees on what it means. And the disagreement isn’t cosmetic. It goes all the way down to the foundations of physics, the nature of reality, and the question of whether the universe is fundamentally made of stuff or made of structure.

I think the emergence of AI systems has made this question newly urgent. Because if you build a system that processes “information” at superhuman scale and speed, and you can’t say what information is, then you can’t say what your system is doing. And if you can’t say what it’s doing, you can’t say whether it understands, whether it knows, or whether its outputs mean anything at all.

Shannon’s Brilliant Dodge

Claude Shannon published “A Mathematical Theory of Communication” in 1948, and it’s one of the most consequential papers in the history of science. Shannon solved a concrete engineering problem: how do you quantify the capacity of a communication channel? How many bits per second can you push through a telephone wire?

His answer was elegant. He defined information mathematically as the reduction of uncertainty. If you flip a fair coin, the outcome carries one bit of information, because it resolves one binary question. If you roll a fair die, the outcome carries about 2.58 bits, because it resolves more uncertainty. The more uncertain you were before the message, and the more that uncertainty is reduced after, the more information the message carries.

Shannon’s genius was in what he left out. His definition of information says nothing about meaning, truth, or reference. The sentence “the cat sat on the mat” and the sentence “the glorb flixed the quaz” carry the same amount of Shannon information if they have the same statistical properties. Shannon himself was explicit about this: “The fundamental problem of communication is that of reproducing at one point either exactly or approximately a message selected at another point. Frequently the messages have meaning; that is they refer to or are correlated according to some system with certain physical or conceptual entities. These semantic aspects of communication are irrelevant to the engineering problem.”

That last sentence is one of the most important methodological moves in the history of ideas. Shannon didn’t solve the problem of meaning. He bracketed it. He said: I’m going to define information in a way that’s useful for engineering, and I’m going to deliberately ignore everything about content, truth, significance, and reference. You can build an entire theory of communication on top of this, and it works beautifully. But it doesn’t tell you what information means. It tells you how much of it there is.

This is the dodge. Shannon information is a measure of quantity without a theory of quality. It tells you the size of the message without telling you what the message says. And for engineering purposes, that’s exactly right. You don’t need to understand the content of a phone call to design a telephone network. You just need to know the bandwidth requirements.

But when we try to use Shannon information as a foundation for understanding AI systems, knowledge, or reality itself, the dodge becomes a problem. Because the questions we’re trying to answer aren’t engineering questions. They’re semantic questions. Does this AI system understand the information it processes? Is there a difference between processing information and comprehending it? Can information exist without a mind to interpret it?

Shannon’s theory is silent on all of these.

Floridi’s Correction: Semantic Information

Luciano Floridi, who essentially founded the modern philosophy of information, spent decades trying to fill the gap Shannon left. His key move was defining semantic information as “well-formed, meaningful, and truthful data.”

Let me unpack each of those terms, because they do a lot of work.

Requirement	What it means	Why it matters
Well-formed	The data follows the rules of its representational system (grammar, syntax, format)	Noise and random bit strings aren’t information, even if they have high Shannon entropy
Meaningful	The data is about something, it has semantic content	Syntactically valid gibberish (“colorless green ideas sleep furiously”) satisfies well-formedness but not meaningfulness
Truthful	The data corresponds to reality	False statements aren’t information, they’re misinformation

That last requirement is the controversial one. Floridi calls it the “veridicality thesis”: genuine information must be true. False information is a contradiction in terms, like “married bachelor.” If someone tells you that Paris is the capital of Germany, they haven’t given you information. They’ve given you misinformation. The word “information” has positive epistemic valence. It implies truth.

This is a strong claim, and not everyone agrees with it. Critics point out that it conflicts with ordinary usage (“the information in this report turned out to be false”) and that it makes the concept of “false information” incoherent. Fred Dretske, another major figure in the philosophy of information, defines information without the truth requirement: for Dretske, information is anything that carries a signal about the state of the world, whether or not that signal is accurate.

The disagreement between Floridi and Dretske matters for AI. Under Floridi’s definition, an LLM that hallucinates doesn’t produce information. It produces well-formed, meaningful data that fails the truth test. Under Dretske’s definition, the LLM does produce information (it carries a signal about the training data distribution), but the signal may be misleading.

When people say “AI processes information,” they’re usually using Shannon’s definition implicitly: the AI processes bit patterns. But the interesting questions, about understanding, knowledge, and meaning, require Floridi’s or Dretske’s richer definitions. And under those richer definitions, it’s genuinely unclear whether AI systems process information at all, or whether they process data that becomes information only when a human interprets it.

The Four Faces of Information

Floridi identifies four distinct phenomena that we lump together under the word “information”:

Information about something. A train timetable contains information about train departures. A weather forecast contains information about atmospheric conditions. This is the most intuitive sense: information as a representation of something external.

Information as something. DNA contains information. Fingerprints contain information. Tree rings contain information. In this sense, information is a natural phenomenon, a pattern in the physical world that can be read by an observer. The information isn’t created by a mind. It’s already there in the structure of the thing itself.

Information for something. An algorithm is information for computing a result. A recipe is information for baking a cake. Instructions, procedures, plans: these are all information in the sense of being specifications for action.

Information in something. A constraint is information. The shape of a key contains information about the shape of the lock. A mathematical equation contains information in its structure, in the relationships between its terms. This is the most abstract sense: information as pattern, as structure, as constraint on possibility.

These four senses are distinct but compatible. A DNA molecule is simultaneously information about an organism’s traits, information as a physical structure, information for building proteins, and information in the specific pattern of nucleotide base pairs.

The reason this taxonomy matters is that different theories of information pick different senses as fundamental. Shannon picks “information about” (signals carrying data about a source). Floridi emphasizes all four but gives primacy to “information about” in its semantic version. Wheeler, as we’ll see, picks “information in” and argues that it’s fundamental to reality itself.

Wheeler’s Wild Idea: It From Bit

In 1989, the physicist John Archibald Wheeler (who coined the terms “black hole” and “wormhole,” so his track record on influential ideas is pretty good) proposed a radical thesis: the universe is fundamentally informational. Physical reality arises from information, not the other way around.

He summarized it in three words: “it from bit.”

“It from bit symbolizes the idea that every item of the physical world has at bottom an immaterial source and explanation; that which we call reality arises in the last analysis from the posing of yes-no questions and the registering of equipment-evoked responses.”

This is a stunning claim. Wheeler isn’t saying that information is a useful way to describe reality (nobody would argue with that). He’s saying that information is what reality is. The electron isn’t a thing that can be described by information. The electron is information. Mass, charge, spin: these are all just information, patterns of yes-no answers to questions posed by measurements.

Wheeler’s idea sounds metaphysical, but it has a physical anchor: quantum mechanics. In quantum theory, a particle doesn’t have definite properties until it’s measured. Before measurement, it exists in a superposition of possible states. The measurement “asks a question” (spin up or spin down? here or there?), and the result is a bit of information. Wheeler argued that this isn’t just a feature of our measurement process. It’s a feature of reality itself. Reality is constituted by the answers to yes-no questions. It from bit.

This connects to an idea from Rolf Landauer, who argued in 1961 that information is physical. Not metaphorically physical. Actually, thermodynamically physical. Erasing one bit of information in a computational device necessarily dissipates at least kT ln 2 joules of energy as heat (where k is Boltzmann’s constant and T is the temperature). This is Landauer’s principle, and it’s been experimentally verified.

Landauer’s principle is philosophically important because it establishes that information isn’t just an abstract mathematical quantity. It has physical consequences. It takes up space (in whatever medium stores it). It requires energy to process. It generates heat when destroyed. It obeys thermodynamic laws.

If Wheeler is right that reality is fundamentally informational, and Landauer is right that information is fundamentally physical, then we get a strange loop: the physical is informational, and the informational is physical. They’re not two different things described by one framework. They’re one thing described by two frameworks.

A 2025 review paper in Entropy (“Landauer’s Principle: Past, Present and Future”) traces how this idea has developed since Landauer’s original proposal. The paper connects Landauer’s principle to Wheeler’s “it from bit” paradigm and argues that the principle bridges information theory and thermodynamics in a way that has become increasingly relevant to quantum computing, where the physical costs of information processing are no longer negligible but central to the engineering.

The Chinese Room, Revisited From an Information Perspective

I wrote about Searle’s Chinese Room in my philosophy of mind post, but it’s worth revisiting from an information-theoretic perspective, because the argument looks different through this lens.

Searle’s claim is that the person in the room manipulates symbols without understanding them. Syntax isn’t semantics. The room processes information (in Shannon’s sense) without comprehending it (in the semantic sense).

But here’s the thing. From an information-theoretic perspective, the Chinese Room is doing something interesting even without understanding. It’s transforming information. It takes an input (Chinese characters), applies a function (the rulebook), and produces an output (Chinese characters). The transformation preserves certain structural relationships between input and output. In fact, it preserves exactly the structural relationships that a Chinese speaker would preserve.

Under structural realism (which I’ll get to shortly), structure is all there is. If the Chinese Room preserves the right structure, then it preserves the right information, and the question of whether it “understands” becomes a question about what understanding adds beyond structural preservation.

Consider an analogy. A prism takes white light (input), applies a physical transformation (refraction), and produces a spectrum (output). Does the prism “understand” optics? Obviously not. But it preserves the structural relationships between wavelength and refraction angle perfectly. It’s a structural information processor, and a good one.

The question is whether the difference between the prism and a human physicist who “understands” optics is a difference in kind (the physicist has something the prism fundamentally lacks) or a difference in degree (the physicist has more complex structural processing, but the same basic type of operation).

Searle says: kind. The physicist has intentionality, meaning, understanding. The prism (and the Chinese Room) have none of these.

The information-theoretic perspective suggests: maybe degree. The physicist’s “understanding” might just be a very complex form of structural information processing, one that includes self-referential loops (thinking about thinking), counterfactual reasoning (what would happen if…), and flexible deployment across novel situations. All of which could, in principle, be captured by a sufficiently complex structural transformation.

This is the view that Floridi calls “informational structural realism”: reality is fundamentally structural, structures are fundamentally informational, and the distinction between “processing information” and “understanding information” might be a distinction without a difference at a deep enough level of analysis.

I don’t fully buy this, for reasons I’ll explain. But it’s a serious position that forces you to say exactly what understanding adds beyond structural preservation. And most people, when pressed, have trouble articulating that.

Structural Realism: What’s Real Is Structure

Let me talk about structural realism, because it’s the ontological position most naturally connected to the philosophy of information, and because it offers a way to think about AI that most AI researchers haven’t encountered.

Structural realism comes in two flavors:

Epistemic structural realism (ESR), associated with John Worrall (1989), says: we can’t know the nature of things in themselves, but we can know their structure. When science changes (from Newton to Einstein, from classical to quantum), the structural content of the old theory is preserved in the new one, even though the ontological furniture changes. Newton’s inverse-square law is structurally preserved in general relativity, even though the underlying ontology changed from “gravitational force” to “spacetime curvature.” What’s preserved across theory change is structure, not stuff.

Ontic structural realism (OSR), associated with James Ladyman, Don Ross, and others (particularly their 2007 book Every Thing Must Go), makes a stronger claim: structure is all there is. There are no “things” underlying the structure. The electrons, quarks, and fields of physics are not objects with properties. They’re nodes in a web of structural relations. What’s real is the pattern, not the stuff that makes the pattern.

Ladyman and Ross assert, bluntly: “the world is not made of anything.” By which they mean: the world is made of structures, patterns, and relations, not of things that have structures as properties. The structure is the thing.

This matters for the philosophy of information because, if OSR is right, then information (understood as structure, pattern, relation) isn’t just a useful way to describe reality. It is reality. Floridi’s informational structural realism connects these ideas explicitly: the ultimate nature of reality is informational, not material. What exists are informational structures, patterns of differences.

And this matters for AI because it changes the question we’re asking. If reality is fundamentally structural/informational, and AI systems process structural/informational content, then the gap between “what AI does” and “what reality is” narrows considerably. The objection that AI “merely” processes information becomes less compelling when information is promoted from “description of reality” to “constitution of reality.”

The Physical Substrate of Information

Let me bring this back to earth with a question that connects physics, philosophy, and computer science: where does information live?

Shannon treated information as abstract. Bits are mathematical objects, defined by probability distributions. But Landauer showed that bits are always physical. Every bit is stored in some physical medium (a transistor, a magnetic domain, a quantum state, a pattern of neural activity). And every physical storage of information is subject to physical laws: thermodynamics, quantum mechanics, the speed of light.

This creates a tension. On one hand, information seems substrate-independent. The same sentence can be stored in ink on paper, as voltage patterns in RAM, as magnetic orientations on a hard drive, or as patterns of neural activation in a brain. The information is “the same” regardless of the substrate. This is the key premise of functionalism in philosophy of mind and computationalism in philosophy of information.

On the other hand, information is always physically realized, and the physical realization matters. The same logical computation, run on different physical substrates, can have different thermodynamic costs, different noise characteristics, different failure modes. Quantum information (stored in quantum states) behaves fundamentally differently from classical information (stored in classical states). Information on a hard drive persists when you turn off the power. Information in RAM doesn’t. Information in a biological brain is plastic, adaptive, and context-sensitive in ways that information on a silicon chip isn’t.

This suggests that the relationship between information and its physical substrate is more intimate than the functionalist picture implies. Information isn’t perfectly substrate-independent. It’s substrate-influenced: the physical medium shapes how the information can be accessed, transformed, combined, and destroyed.

For AI, this has a concrete implication. The “same” information, processed by a biological neural network and by an artificial neural network, is processed differently in ways that might matter for understanding, consciousness, and meaning, even if the functional input-output behavior is identical. If Landauer is right that information is physical, and if the physical substrate shapes the information processing in non-trivial ways, then you can’t assume that functional equivalence implies semantic equivalence.

Syntax, Semantics, and the Grounding Problem

Here’s the deepest tension in the philosophy of information, as it relates to AI: the relationship between syntax and semantics.

Syntax is about form: the rules that govern how symbols can be combined. “The cat sat on the mat” is syntactically valid English. “Cat the on mat sat the” is not.

Semantics is about meaning: what the symbols refer to, what they’re about. “The cat sat on the mat” means something. It refers to a cat, a mat, and a spatial relationship between them.

Searle’s thesis was that computers have syntax but not semantics. They can manipulate symbols according to formal rules, but they can’t assign meaning to those symbols. Meaning requires something that computation alone can’t provide: a connection between the symbol and the thing it represents. This connection is what philosophers call “grounding.”

The symbol grounding problem, articulated by Stevan Harnad in 1990, asks: how do symbols get their meaning? If you look up a word in a dictionary, the definition is in terms of other words. Those words have definitions in terms of yet other words. At some point, the chain of definitions must be grounded in something that isn’t a symbol: in direct sensory experience, in physical interaction with the world, in some non-linguistic connection between the representation and the thing represented.

This is where LLMs face their most serious philosophical challenge. An LLM learns everything from text. Its “understanding” of the word “cat” is constituted entirely by the statistical relationships between “cat” and every other word in its training corpus. It has never seen a cat, touched a cat, heard a cat purr, or been scratched by a cat. Its representation of “cat” is grounded in language, which is grounded in more language, all the way down. There’s no perceptual floor, no experiential bedrock, no place where the symbol chain terminates in the world itself.

Is this a problem? It depends on your theory of meaning.

If meaning requires direct causal connection to referents (as Harnad argues), then LLMs don’t have it, and their outputs are semantically empty, no matter how syntactically sophisticated.

If meaning is constituted by functional role (the network of inferential relationships a symbol participates in), then LLMs might have it, because their internal representations participate in enormously rich inferential networks. The model’s representation of “cat” is connected to “mammal,” “pet,” “whiskers,” “purring,” “scratching,” “litter box,” and thousands of other concepts in ways that support flexible reasoning about cats.

If meaning is structural (what matters is whether the relationships between representations mirror the relationships between things in the world), then LLMs partially have it. The geometric relationships in LLM embedding spaces do mirror many real-world relationships. But they also encode training data biases, statistical artifacts, and purely linguistic regularities that don’t correspond to anything in reality.

The honest answer is that we don’t have a good enough theory of meaning to settle this question. We know what meaning isn’t (it isn’t just syntax). We have candidates for what it might be (causal grounding, functional role, structural correspondence). But we can’t definitively say which candidate is right, which means we can’t definitively say whether LLMs are semantically empty or semantically rich.

Information Processing vs. Information Understanding

Let me try to make the distinction between processing and understanding as concrete as I can, because this is where the ontology of information meets the epistemology of AI.

A calculator processes information. It takes numeric inputs, applies mathematical operations, and produces numeric outputs. Nobody thinks a calculator understands arithmetic. The processing is entirely syntactic: symbol manipulation according to formal rules, with no semantic content from the calculator’s perspective.

A human mathematician processes information too. They take mathematical statements, apply transformations, and produce new mathematical statements. But (we assume) they also understand what they’re doing. They know why the transformations are valid, what the statements mean, and when the results are significant.

The question is: what’s the difference? What does the mathematician have that the calculator lacks?

Candidates for the answer include:

Candidate	What it means	Problem
Consciousness	The mathematician has subjective experience of doing math	Explains understanding by appealing to something equally mysterious
Intentionality	The mathematician’s mental states are about mathematical objects	Circular if intentionality itself requires understanding
Causal grounding	The mathematician’s symbols are causally connected to the world through experience	Math is about abstract objects, not physical ones, so what are mathematical symbols grounded in?
Flexible deployment	The mathematician can apply their knowledge in novel situations	LLMs can also apply knowledge flexibly in novel situations, at least sometimes
Meta-cognition	The mathematician knows what they know and what they don’t know	True, but meta-cognition might be achievable in AI systems

None of these candidates is fully satisfying. Each one either appeals to something equally mysterious (consciousness, intentionality) or identifies a feature that AI systems already partially possess (flexible deployment, possibly meta-cognition).

This is the hard problem of information, and it’s analogous to the hard problem of consciousness. We can describe all the functional aspects of understanding (flexibility, meta-cognition, appropriate response to novel situations), and even if we build a system that has all of those functional aspects, we still can’t be sure it “understands” in the way a human does. There might be something more to understanding, something beyond function, and we can’t specify what it is.

Or there might not be. Maybe understanding just is sufficiently complex, flexible, meta-cognitive information processing. Maybe there’s no further fact of the matter. Maybe the question “does the AI really understand?” is like asking “is this number really prime?” The question has a definite answer, and the answer depends entirely on the structural properties of the thing in question, not on some additional metaphysical ingredient.

The AI Mirror: What Building Systems Taught Us About Information

The story of the philosophy of information is the story of a concept getting more complicated the harder you look at it. Shannon started with a clean, mathematical definition. Floridi added semantic content. Wheeler promoted information to the fundamental fabric of reality. Landauer anchored it in physics. The structural realists absorbed it into their ontology. And AI forced everyone to confront the question they’d been avoiding: is information processing the same thing as information understanding?

What AI systems have taught us, whether or not we wanted to learn it:

1. Shannon information is necessary but not sufficient.

You need a theory of quantity (how much information?) to build communication systems. But you also need a theory of quality (what does the information mean? is it true? is it relevant?) to build systems that interact with humans in meaningful ways. Shannon gave us quantity. Nobody has given us quality in a way that’s as mathematically rigorous, and that gap is the source of most philosophical confusion about AI.

2. The syntax/semantics distinction might not be a binary.

Searle treated it as a binary: you either have semantics or you don’t. But the evidence from LLMs suggests a continuum. These systems have richer “understanding” than a lookup table, poorer understanding than a human expert, and something in between that we don’t have good vocabulary for. Maybe semantics isn’t all-or-nothing. Maybe it comes in degrees, and the question isn’t “does the AI have semantics?” but “how much semantics does the AI have?”

3. Grounding might be achievable through language alone.

This is the most controversial claim, and I’m not sure it’s right. But it’s worth taking seriously. If meaning is constituted by the totality of inferential relationships a concept participates in (as inferentialist philosophers like Robert Brandom argue), then a system trained on all of human language might have acquired meanings for its symbols, because it has acquired the inferential relationships. The meanings would be different from ours (because ours are also grounded in perception and action), but they might be genuine meanings nonetheless, not zero semantics but a different kind of semantics.

4. Information is relational, not intrinsic.

A bit pattern in a hard drive isn’t intrinsically “about” anything. It becomes information about something only in relation to an interpreter, a context, a purpose. This is true for human language too. The word “bank” means “financial institution” or “edge of a river” depending on context. But humans carry their interpretive context with them (in their knowledge, their experience, their goals). AI systems carry interpretive context too (in their training, their architecture, their prompt), but it’s a different kind of context, and whether it’s the right kind is the open question.

5. The ontology of information constrains the epistemology of AI.

What you think information is determines what you think AI does. If information is just bit patterns (Shannon), then AI processes information, full stop. If information is meaningful truth (Floridi), then AI might not process information at all, because its outputs aren’t guaranteed to be true. If information is the fabric of reality (Wheeler), then AI is doing something cosmologically significant, even if it doesn’t know it. Your ontology precedes your assessment.

What Kind of Thing Is Information, Really?

After spending this much time with the question, I think the honest answer is: information is not one kind of thing. It’s a family of related concepts unified by a metaphor.

The metaphor is: something is transmitted, preserved, or transformed. A signal carries information from sender to receiver. DNA carries information from parent to offspring. A book carries information from author to reader. An LLM carries information from training corpus to output.

But the nature of what’s “carried” differs in each case. Shannon information is a statistical quantity. Floridi’s semantic information is a propositional content. Wheeler’s “it from bit” information is a feature of physical reality. Biological information is a functional template for molecular assembly. The word “information” papers over these differences, creating the illusion of a unified concept where there are actually several related but distinct ideas.

This matters because many arguments about AI exploit the ambiguity. “AI processes information” is true under Shannon’s definition. “Therefore AI understands information” is a non sequitur, because understanding requires semantic information, not just Shannon information. The argument works only if you equivocate between the two senses.

Similarly, “information is physical” (Landauer) doesn’t imply “physical processing of information constitutes understanding” (functionalism). The physical reality of information is about thermodynamics and energy costs. Understanding is about meaning and reference. These are different questions that happen to use the same word.

A Taxonomy of Gaps

Let me try to organize the conceptual landscape. The philosophy of information, as applied to AI, revolves around several gaps, each representing an unresolved philosophical question.

Gap	Between	Status
The Shannon-Floridi gap	Quantitative information (how many bits?) and qualitative information (what does it mean?)	Open. No rigorous bridge.
The syntax-semantics gap	Formal symbol manipulation and meaningful representation	Possibly a continuum rather than a gap, but the endpoints are clear.
The processing-understanding gap	Computational transformation of information and genuine comprehension	The central question. No consensus.
The data-information gap	Raw signals and organized, contextual, meaningful content	Relatively well understood. Context and organization are key.
The information-knowledge gap	Propositional content and justified, integrated, action-guiding belief	Well explored in epistemology. Knowledge requires more than information.
The physical-abstract gap	Information as physical state (Landauer) and information as mathematical quantity (Shannon)	Bridged by Landauer’s principle, but philosophical implications still debated.

These gaps aren’t bugs in our understanding. They’re features of the concept. Information sits at the intersection of physics, mathematics, semantics, and cognition, and the concept means something different in each domain. The challenge is to maintain clarity about which sense you’re using, and to resist the temptation to slide between senses when constructing arguments about AI.

The Information Turn in Philosophy

There’s a broader context here that’s worth noting. Over the past two decades, philosophy has undergone what Floridi calls “the information turn”: a reorientation of fundamental questions around the concept of information.

Metaphysics asks: what exists? The informational answer (Wheeler, Floridi): what exists is informational structure.

Epistemology asks: what can we know? The informational answer: we can know the informational structure of things, not their intrinsic nature (epistemic structural realism).

Ethics asks: how should we act? The informational answer (Floridi’s information ethics): we should minimize the destruction of informational objects and maximize the flourishing of the “infosphere.”

Philosophy of mind asks: what is a mind? The informational answer: a mind is a sufficiently complex information-processing system that generates semantic content.

Philosophy of science asks: what does science discover? The informational answer: science discovers structural/informational relationships between phenomena.

If the information turn is right (or even partly right), then AI isn’t just a technology. It’s a philosophical instrument. It’s a machine that forces us to be precise about concepts (information, meaning, understanding, knowledge) that we’ve been using loosely for centuries. And being forced to be precise reveals that we understand these concepts much less well than we thought.

Where This Leaves Us

The ontology of information is messy. The concept refuses to reduce to a single clean definition. It’s simultaneously a mathematical quantity, a semantic content, a physical state, a structural relation, and (possibly) the fabric of reality itself. These senses are related but not identical, and arguments that conflate them produce confusion.

For AI specifically, the takeaway is that most debates about what AI systems “do with information” are actually debates about what information is. People who say “LLMs just process statistical patterns, there’s no understanding” are using a Shannon-type definition. People who say “LLMs have learned the structure of the world” are using a structural-realist definition. People who say “LLMs don’t really know anything because they can produce false outputs” are using a Floridi-type definition. They’re all correct, given their definitions. They’re all talking past each other.

The philosophy of information doesn’t resolve these debates. What it does is make the source of disagreement visible. It shows you that the question “does AI understand information?” can’t be answered until you answer a prior question: “what kind of thing do you think information is?”

And that question, despite Shannon’s elegant math and Floridi’s careful philosophy and Wheeler’s cosmic vision, remains genuinely open. We built machines that process information at scales our ancestors couldn’t imagine. We still don’t know what information is. There’s something important in that gap, something about the relationship between building and understanding, between engineering and philosophy, between making things work and knowing why they work.

We may have to settle for the humbling possibility that information, like consciousness, like knowledge, like meaning itself, is a concept we can use effectively without fully understanding. Our machines process it. Our theories describe it. Our philosophy circles it.

But we haven’t caught it yet.