Why Definitions Matter

Most AI projects fail not because the technology does not work, but because no one agreed on what the system was supposed to mean. Teams build pipelines, train models, and deploy agents while quietly disagreeing about what a "customer" is, what "current" means, or whether a transaction that was reversed still counts. The system runs. The results are wrong. And no one can explain why, because the meaning was never made explicit.

This article offers working definitions of three terms that get conflated constantly in enterprise AI: data, information, and knowledge. More importantly, it explains what I mean by semantics, because that word in particular has been stretched to cover so many things that it has stopped being useful.

These definitions draw on three sources: Charles Sanders Peirce's semiotics, Ludwig Wittgenstein's philosophy of language, and Ronald Stamper's semiotic ladder as adapted by Jan Dietz in his work on Enterprise Ontology. The goal is not to settle philosophical debates. The goal is to give architects, AI designers, and product teams a vocabulary precise enough to build on.

These are working definitions. I find them useful for thinking clearly about information and knowledge-based systems.

The Semiotic Ladder: How Meaning Emerges in Layers

Ronald Stamper developed the semiotic ladder as part of his work on Organizational Semiotics, extending the classical syntax/semantics/pragmatics division from Charles Morris. Jan Dietz further adapted this ladder for Enterprise Ontology and DEMO, adding the physical and social layers. It offers a useful way to understand how meaning builds up in layers. Each level depends on the one below it. If you try to skip levels, you end up assuming things that were never actually provided. Each level contributes something the lower levels simply do not have.

Physical The material substrate: electrical impulses, sound waves, marks on paper. No informational structure and no meaning, just physical substance and causal interaction.
Empiric Encoding appears. How are signals expressed? Roman letters, Morse code, binary digits. The patterns used to represent signals, independent of what they mean.
Syntactic Structure and rules emerge. Grammar, protocols, data formats. We can distinguish well-formed expressions from malformed ones, but still have no meaning, only formalism.
Semantic Meaning enters. Signs become associated with concepts. Symbols refer to things, whether real or modeled. This is where most discussions of "semantic technology" claim to operate.
Pragmatic Intent appears. Why is this being said? What is the purpose? What action or commitment does it imply? Meaning becomes situated in goals and contexts.
Social Shared conventions, norms, roles, and mutual commitments. Meaning is now embedded in a community of practice, not just an individual interpretation.

Dietz groups these levels into three categories. The lower levels (empirics and syntax) concern form: how signs are encoded and structured. The middle levels (semantics and pragmatics) concern content: what signs mean and why they are used. The top level (social) concerns commitment: the shared agreements that make communication binding. Dietz calls these forma, informa, and performa.

This ladder makes something clear: you cannot reach knowledge by operating only at the semantic level. Semantics gives you meaning, but meaning without intent and social context is inert. It does not act. It does not commit. It does not know.

This is why simplistic equations like "knowledge = data + semantics" fail. They skip levels. I will propose an alternative formulation after explaining what I mean by semantics.

Data: Signals Before Interpretation

Data sits at the bottom of our ladder. It is raw symbols without structured meaning: the bytes, the signals, the uninterpreted marks.

Data is necessary but not interesting on its own. It becomes something more only when we impose structure (moving to syntax) and interpretation (moving to semantics). The mistake many systems make is treating data as if it were already meaningful, as if collecting and storing something automatically makes it significant.

Data is what you have before you have made any conceptual commitments. It is the material from which meaning will be constructed, not meaning itself.

Semantics: Conceptual Distinctions, Not Closed Truth

Here is where clarity matters most.

Semantics is the disciplined articulation of conceptual distinctions, expressed in a conceptual schema, that supports shared reference, reasoning, and the possibility of coordinated action. Its validity emerges through use, correction, and contextual interpretation, not through prior closure.

Two philosophical perspectives inform this definition.

Peirce: Meaning Is Triadic

Charles Sanders Peirce argued that meaning is not a simple two-way relationship between a word and a thing. It is triadic:

The interpretant is the key element. Meaning does not exist in the sign alone, nor in the object alone. Meaning exists in the relationship between sign, object, and interpretation. Without an interpreter, human or otherwise, there is no meaning.

This has direct consequences for so-called "semantic systems." An OWL ontology or RDF graph contains signs and structural relationships. But the interpretant, the meaning, arises only when something interprets those signs in a context of use. The system itself does not "understand" its contents any more than a dictionary understands the words it contains.

Peirce also insisted that referents can be "real or modeled." Not everything a sign refers to must exist in the physical world. Some referents exist because we model them: legal entities, social roles, abstract categories. Their meaning stabilizes through use, not through correspondence to physical objects.

This is why I prefer the term conceptual schema over ontology. Ontology, in its philosophical sense, asks "what exists?" and tends toward fixed, truth-asserting commitments. A conceptual schema asks "what distinctions matter for our purposes?" and allows for provisional, revisable, purpose-relative modeling.

Wittgenstein: Meaning Is Use

Ludwig Wittgenstein argued that, for most cases, the meaning of a word is best understood as its use in the language. Words do not carry meaning like containers. They acquire meaning through participation in structured activities, what he called "language games."

This has practical implications. Meaning is not fixed at design time. A conceptual schema is not validated when you draw it or formalize it. It is validated when:

Semantics is therefore intended at modeling time, where we aim for certain meanings, but tested at runtime, where we discover whether those meanings actually hold up in practice. This directly contradicts any demand for "deterministic semantic termination," the idea that meaning should be closed, complete, and self-executing.

This does not deny the value of formal semantics. It situates formal semantics within practice rather than above it.

What Semantics Is Not

Given this understanding, we can state precisely what semantics is not:

Information: Semantics Without Intent

If semantics is meaning, then information is structured meaning that can be communicated, stored, and reasoned over, but that lacks the pragmatic dimension of purpose.

Information tells you what. It does not tell you why or what to do.

Consider the statement: "Person X has birth date January 15, 1990." This is information. It has semantic content. We know what "Person," "birth date," and "January 15, 1990" refer to. But the information alone does not tell us why we are tracking this, what we should do with it, or what commitments follow from it.

The same information might serve multiple purposes:

The semantics remain the same. The intent changes.

The same is true of axioms, constraints, and business rules. They add structure and discipline to meaning, and they can be evaluated or enforced, but they still do not supply purpose. A rule can be correct and executable and still not be knowledge.

Information is semantics without commitment or intent. Execution does not change that.

Knowledge: Semantics + Intent + Social Context

Knowledge adds what information lacks: purpose, commitment, and grounding in practice.

Knowledge is not merely structured meaning. It is meaning that has been:

  1. Connected to intent. Knowledge serves goals. It exists within a context of action and decision.
  2. Validated through use. Knowledge is tested after the fact. Did it work? Did it enable good decisions? Did it survive challenge?
  3. Embedded in social practice. Knowledge is shared, maintained, and evolved by communities. It carries weight about what matters, not just what is true.

This is why knowledge requires the full semiotic ladder. You cannot have knowledge operating only at the semantic level. You must also engage the pragmatic level (intent, goals, purpose) and the social level (norms, roles, shared commitments).

Knowledge is what happens when semantics becomes actionable within a community of practice.

Now we can state the formula properly:

Knowledge = (Symbols + Structure + Shared Meaning) × Intention

The multiplication sign matters. Without intention, the other components remain latent, present but not activated into something we would recognize as knowledge. That is also the bridge to the governance question taken up in Intent-Driven AI Delegates, where intention is treated not as a slogan but as an explicit architectural constraint.

Peirce and Wittgenstein converge here. For Peirce, the interpretant of a knowledge system is operational: knowledge systems are validated by what they enable, by decisions, actions, and coordinations. For Wittgenstein, understanding is not a mental state prior to action but something that shows up in action. To know is to be able to go on, to use the knowledge appropriately in new situations.

Why This Matters for AI

These distinctions have direct consequences for how we design AI systems, and specifically for where we put LLMs in the architecture.

Most systems marketed as "semantic AI" are operating at the structural or logical level. They provide formal syntax, consistency checking, and inference over asserted propositions. Those are valuable capabilities. But they are not semantics in the sense described here. They lack the interpretant. They lack validation through use. They lack the connection to intent. Calling them semantic creates false expectations about what they can guarantee and who is responsible when they get it wrong.

LLMs fit into a well-grounded AI architecture as interpreters, not authorities. They are good at mapping ambiguous natural language to known conceptual elements, surfacing candidate meanings, and flagging missing context. That is genuinely useful. But an LLM does not define what terms mean in your domain. It does not enforce constraints. It does not decide who has authority to act. And it cannot be held accountable for outcomes.

The missing layers in most AI system designs are not technical. They are: intent, which defines what purpose the AI is actually serving; accountability, which defines who is responsible when the AI acts; and human judgment, which defines what cannot be delegated at all. A conceptual schema, no matter how rigorous, cannot answer those questions. Neither can an LLM. They require explicit modeling of intent, delegation, and governance, which is the subject of the companion article on Intent-Driven AI Delegates.

Conclusion: Semantic Honesty as Practice

Semantics is not a property you achieve once at design time. It is a practice you maintain as the system runs, as the domain changes, and as use reveals gaps that modeling could not anticipate.

The semiotic ladder makes the stakes clear. Data becomes information when it is structured. Information becomes actionable when it is connected to intent. Action becomes responsible when it is embedded in a social context with clear lines of accountability. Skip any of those steps and you have a system that runs but does not know what it is doing or why.

For AI systems specifically, this means three things. First, stop calling systems semantic when they are really syntactic. The label raises expectations the system cannot meet. Second, make intent explicit. It does not emerge from data on its own, and burying it in code makes it invisible to governance and audit. Third, keep human accountability in place, not because AI is untrustworthy, but because accountability is a social act that only persons can perform.

These are not abstract principles. They are design constraints. Applied consistently, they produce AI systems that are auditable, governable, and resilient to the domain change that every enterprise eventually faces. The companion article on Intent-Driven AI Delegates shows what that looks like in practice, and the prototype work at knowledge-foundation.ai shows what it looks like running.

G. Sawatzky works at the intersection of formal conceptual modeling and AI systems. For more on Intent-Driven AI Delegates, see the companion article Intent-Driven AI Delegates: A Governance Framework for AI Systems.