Meaning, Ontology, and the Limits of Fixed Definitions

WhatWhat does this word mean?" sounds like the simplest question in the world, and often it is answered by looking the word up. Sometimes that answer settles the matter. Sometimes it starts an argument. Someone says a system is "not really AI," someone else says it obviously is, and both reach for a definition that the other does not accept. This page looks at why that happens, and at what the word ontology has to do with it, since the word turns up in these disputes more often than it should.

The word comes from two Greek roots: onto-, meaning being or existence, and -logia, meaning study, account, or theory. Taken literally, an ontology is an account of what exists. The broad question behind it is simple to state: what things exist, what kinds of things are there, and how are they related? The trouble is that the word has taken on several substantially different meanings across disciplines. A philosopher, a social researcher, and a knowledge engineer can each use the word correctly while doing quite different things with it. The sections below separate those uses, then show how meaning, context, and definition interact, and finish with a practical way to tell whether a terminology dispute is really about the world, about words, or about which framework is in play.onto: being logia: study

1. Three Disciplines, One Word FoundationalKnowledge that endures for decades — core principles

InIn philosophy, ontology is the branch of metaphysics that investigates the fundamental categories of being: what kinds of things there are, and in what sense each kind exists. In social research, the word refers to a researcher's assumptions about the nature of social reality: whether institutions, categories, and roles are stable features of the world that can be measured, or processes continually produced through interaction. In computer and information science, an ontology is an explicit, formal model of concepts and relations within a defined domain, built so that a machine can use it.the root is 'to be', not 'to exist'

These uses share a concern with categories, entities, properties, and relations. They differ in purpose and in what they are trying to achieve. A philosophical ontology is trying to say what exists. A social ontology is trying to say what kind of reality a research design assumes. A computational ontology is trying to say which distinctions a system will recognise, and how it will reason with them. None of the three is the correct meaning of the word, and the rest of this page moves between them, marking each shift.

Three senses of ontology, compared

Sense	Central question	What it is trying to get right	Typical output
Philosophical (metaphysics)	What kinds of things exist, and in what sense?	Fundamental categories and their relations	An argued position on categories such as objects, properties, events, or persons
Social research	What kind of social reality does this study assume?	Consistency between assumptions about social reality and the methods used to study it	A stated research stance (for example positivist or interpretivist), and the methods it licenses
Computational (formal)	Which entities and relations does a system represent, and what follows from them?	A usable, checkable model for a particular domain and purpose	A set of classes, properties, relations, and constraints, expressed in a formal language

2. What Exists? Philosophical Ontology FoundationalKnowledge that endures for decades — core principles

PhilosophicalPhilosophical ontology sits within metaphysics, and its questions can look abstract until a concrete case is put in front of them. Do physical objects, events, numbers, institutions, and fictional characters exist in the same sense, or in different senses? Is "redness" something that exists, or only a name applied to similar experiences? Is a person an enduring entity, a changing biological process, or a socially sustained identity? Are time and causation features of reality itself, or structures through which people organise their experience? And what distinguishes an entity from its properties and relations in the first place?

Three broad orientations recur in answers to these questions. A realist holds that some kinds of entities, structures, or properties exist independently of whether anyone recognises or describes them. A nominalist holds that general categories do not exist as independent universals: a term such as tree, intelligence, or redness groups particular things together for linguistic or practical purposes. A constructivist (or conceptualist) holds that categories depend, at least partly, on human conceptual activity, social practice, or interpretive frameworks.

These are not a simple binary, and a single thinker can hold different positions for different domains. Someone might be realist about physical objects, constructivist about institutional categories such as "qualification", and nominalist about certain abstractions. Treating the three orientations as a menu, rather than as opposing camps, makes it easier to say exactly which commitment is in play in a given argument.

The key point is that philosophical ontology asks what our categories commit us to believing exists. The philosopher W. V. Quine put this in a compact form: a theory is committed to whatever falls in the range of its variables, so that "to be is to be the value of a variable"[3]. The question it poses is not which word we should prefer. It is what we are signing up for when we adopt a particular vocabulary.variables are placeholders in logic, not just code

3. Social Ontology and Research Method Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

SocialSocial ontology asks what kinds of social entities and processes exist, and how they acquire their stability. Nations, organisations, money, laws, universities, markets, social classes, professional roles, and institutional authority are all examples. A university has buildings and employees, but it cannot be reduced straightforwardly to either. It also consists of rules, roles, expectations, records, practices, relationships, and collective recognition, and many of them depend on collective acknowledgement for the kinds of things they are. John Searle's account of "institutional facts" is a well-known attempt to explain how a fact can be real and still depend on collective acknowledgement[6].

Social researchers tend to take one of two broad orientations. An objectivist or positivist orientation treats social structures as stable enough, and external enough to individuals, that they can be investigated as objects with measurable properties. A constructivist or interpretivist orientation treats social reality as continuously produced and negotiated through language, practice, institutions, beliefs, and interaction. Each orientation licenses different methods, and choosing one is a commitment about the world before any data is collected.the map is not the territory

"Socially constructed" does not mean imaginary or inconsequential. Money, laws, qualifications, and borders are all socially constituted, and all of them have powerful material effects. Ian Hacking's study of what is meant when people say something is socially constructed makes the point that the phrase can attach to very different kinds of claim, and that each needs its own argument[7].

It helps to keep four separate things apart, because the word "ontology" is often used as a loose synonym for all of them:

Ontology

What is the nature of the reality being studied?

Epistemology

What can be known about that reality, and how can knowledge of it be justified?

Methodology

What forms of inquiry follow from those commitments?

Methods

What specific procedures are used to gather or analyse evidence?

Keeping these apart prevents ontology from becoming a vague name for "worldview", "opinion", or "method", which are each a different thing.

4. Formal Ontologies in Computing Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe engineering sense of ontology is an explicit, formal representation of concepts and relationships within a defined domain. The field took its present shape in the late 1990s, through the FOIS conference series that began in 1998[5], and the definition most often quoted in it describes an ontology as "an explicit specification of a conceptualization"[4]. A conceptualization is the informal way a community already carves a domain into kinds of things. An ontology makes that carving explicit enough that a program can act on it.

A formal ontology typically contains:

• Classes (concepts), such as Person, Vehicle, or MedicalCondition;
• Instances, such as one particular person or one particular vehicle;
• Properties, such as hasAge or engineType;
• Relations, such as owns, causes, or isPartOf;
• Constraints and axioms, which determine which combinations are permitted;
• Inference rules, which allow new statements to be derived from existing ones.

A small example makes this concrete:

Vehicle
  ├── Car
  ├── Bicycle
  └── Train

Person ──owns──> Vehicle
Vehicle ──hasIdentifier──> Identifier

Rules:
  Every Car is a Vehicle.
  No Bicycle is a Car.
  A registered road vehicle has a registration identifier.

This model does not settle the metaphysical nature of vehicles, ownership, or identity. What it does is provide a usable representation for one particular system, with the distinctions that system needs and no others. Standards such as RDF and the Web Ontology Language (OWL) make this kind of representation shareable across systems, and they underpin linked data, knowledge graphs, biomedical vocabularies, and interoperability between databases[8].

A formal ontology is not necessarily a complete account of reality, a neutral representation, a universal classification, a simple taxonomy, or a dictionary. A taxonomy mainly arranges classes into broader and narrower categories. A formal ontology can also express properties, constraints, relations, identity conditions, and logical commitments, which is what gives it the ability to reason. For a fuller account of how these pieces work in practice, the site's informatics treatment of ontologies and frames develops the software side in detail.cf. frames: slots vs logic

The key point is that a computational ontology externalises distinctions that human participants would otherwise infer from background knowledge. It supplies context to a machine, but only the context its designers chose to represent and managed to represent correctly.

5. Meaning Is Relational, Not Contained in the Word FoundationalKnowledge that endures for decades — core principles

AA word does not carry a complete, context-independent meaning, like a label fixed permanently to an object. Meaning emerges through a network of relations involving signs, possible referents, interpretive habits, situations, and purposes. The American philosopher Charles Sanders Peirce built one of the most influential accounts of this, and his starting point is a three-way relation rather than a simple pairing of word and thing.

Text description of the diagram

A triangle with three corners. The top corner is the interpretant. The bottom left corner is the sign, and the bottom right corner is the object. Each corner is joined to the other two. A dashed frame labelled "context, practice, history" surrounds the whole triangle, so that context shapes the entire three-way relation.

The three elements are these. The sign is something that functions as a sign: a word, image, gesture, sound, or data value. The object is that to which the sign refers or directs attention. The interpretant is the understanding, effect, or further sign produced by interpreting the original sign. Peirce's own wording is that a sign "stands to somebody for something in some respect or capacity," and that it creates in the mind of the interpreter "an equivalent sign, or perhaps a more developed sign"[1].

Deeper note: what the interpretant can be

The interpretant is easy to misread as a private mental picture. Peirce's own definition makes it a sign, and in practice it can take many forms: a concept, a disposition to act, a translation into another language, a response, a rule of interpretation, or a further sign. It is the effect of the sign on an interpreter, expressed as something that can itself be interpreted, which is why Peirce treats the process as one that can keep going.

Meaning, on this account, is produced through semiosis, an ongoing process in which signs generate interpretants, and those interpretants can in turn become signs. The important feature is the irreducible three-way relation. A common way of drawing the account as a single arrow, from sign to object to interpretant, suggests a temporal sequence that Peirce did not intend. The triad is a set of relations that hold at once, not a pipeline.not a pipeline but a simultaneous triad

6. Layers of Context Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe word run shows how context constrains interpretation. In software, to run means to execute a program. In baseball, a run is a unit of scoring. In finance, a run can mean a sustained movement in a price, or a rush to withdraw assets. In everyday speech, it means moving quickly on foot. Each reading is available, and the one selected depends heavily on what the participants are doing. Context in this sense is the activity in which the word is used, and it is often what decides the meaning before any dictionary is consulted.

Several distinct layers of context do this work:

Pragmatic context

The goals and activities of the participants. The same word selects different meanings in a programming session, a sports broadcast, and a trading floor.

Indexical context

Expressions such as I, here, now, this, and we depend on the circumstances of utterance. A system that stores "the meeting is here tomorrow" without recording the speaker, the location, and the date has lost the information needed to resolve the statement at all.

Systemic or contrastive context

Terms take part of their meaning from their relations and contrasts within a larger set. Warm is partly understood through its relation to hot, cool, and cold, so changing the surrounding categories can move the practical boundary of the term.

Frame or schema context

A word can activate structured background knowledge. Buyer implies a seller, goods or services, an exchange, a price, a transfer of ownership, and obligations between the parties.

Historical and communal context

Meanings change over time and differ between communities. A technical community may use a familiar word in a specialised way, while public, commercial, and academic usage diverge.

The indexical and frame layers are where formal systems most often go wrong. Software cannot safely be assumed to share the background knowledge a human participant brings, so frames that people take for granted often need to be stated. The same logic applies to time. Properties such as validFrom, validUntil, assertedAt, supersedes, and appliesInContext let a system limit a claim to the period and setting in which it holds, rather than letting an old statement pass silently as current.indexicals: words like 'here' or 'I' that point to context

7. Use, Language Games, and Kinds of Definition Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

Wittgenstein'sWittgenstein's best-known slogan is that "the meaning of a word is its use in the language"[2]. The slogan is often flattened into the claim that words can mean whatever anyone wants them to mean. The stronger and more defensible point is that meaning is stabilised through publicly recognisable practices, conventions, and ways of living together, and that the stability is real even when it is not total.

Wittgenstein's term for this is the language game. The word model works differently in fashion, in science, in mathematics, in engineering, and in machine learning. None of those uses is automatically the uniquely correct essence of the word. Each is made intelligible by the practice in which it operates, and a dispute between them is often a dispute about which practice is the right setting for a given question.

Definitions remain useful inside these games, but they do different jobs, and it helps to ask which kind of definition is on the table. A definition may be:

Lexical

describing established usage, as a dictionary entry does;

Stipulative

assigning a meaning for the purposes of a particular discussion;

Operational

specifying how something will be measured or detected;

Technical

standardising a term within a professional community;

Persuasive

presenting a contested position as if it were neutral.

A definition's kind shapes what it can legitimately do. A lexical definition can tell you how a word is usually used. It cannot settle how a measurement should be taken, and a persuasive definition is often most dangerous when it is mistaken for a lexical one.lexical: reports usage persuasive: argues for a use

8. Definitions Are Instruments, Not Hidden Essences Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TheThe claim that "words have definitions" is true, and it can be made more useful by saying what definitions actually do. Definitions can reduce ambiguity, coordinate a community, establish legal or technical boundaries, support measurement, permit consistent implementation, and make disagreement visible. That last function is easy to undervalue. A definition that is explicit is one that someone can disagree with precisely.

The claim does not establish several things by itself. A definition does not prove that a category corresponds to a natural division in reality. It does not show that one community's definition governs every context, that category boundaries are timeless, or that borderline cases cannot exist. It does not mean a technical definition overrides ordinary usage, and it does not settle the metaphysical status of whatever the word names.

A useful way to hold these together is this: definitions do not merely report reality, and they do not create reality without constraint. They organise distinctions for particular forms of inquiry and action. Dictionaries are mostly descriptive, recording how words are used. Technical standards, laws, and controlled vocabularies, by contrast, may deliberately prescribe usage within a defined context. Both are legitimate, and confusing one with the other is a common source of argument.

9. Case Study: AI, LLM, Chatbot, Agent Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

TermsTerms like AI, LLM, chatbot, and agent are a good test of everything above, because they are used at several levels of abstraction at once. The two structures below are not the same tree, and they overlap rather than forming one clean hierarchy. The first describes where a model sits among computational approaches. The second describes the parts of a system built around a model. Expand each to see how the levels relate.

Structure 1: where language models sit among approaches to AI

• Artificial intelligence
  • Symbolic and rule-based approaches
  • Search and planning
  • Machine learning
    • Deep learning
      • Large language models
  • Hybrid systems

Structure 2: the parts of a conversational AI system

• Conversational AI system
  • One or more language models
  • Retrieval
  • Memory or state
  • Tools and external services
  • Orchestration
  • Safety and policy mechanisms
  • User interface

Read against each other, the two structures show the category relations. Saying that an LLM is a kind of AI places it within a broad disciplinary and engineering category, and that is simply correct. A large language model is a type of computational model, most often transformer-based[9]. A chatbot is a system or application organised around conversational interaction. A chatbot may use an LLM, but it may also include retrieval, tools, memory, routing, policy checks, and other models. Some chatbots use no LLM at all, and a scripted interface that follows fixed branches is still a chatbot in ordinary usage.

Three ways of saying "that chatbot is an LLM"

• As shorthand, it is acceptable when attention is focused on the model itself, for example when comparing the model's accuracy on a benchmark.
• As a description of the whole system, it is technically incomplete, because the behaviour a user sees may depend on retrieval, tool calls, and orchestration that the model does not contain.
• As an explanation of system behaviour, it can be misleading when the behaviour depends heavily on those non-model components.

Two simplifications are worth avoiding. An LLM is not adequately described as only a "statistical next-token predictor". Next-token prediction is a central mechanism in training and generation for many LLMs, but deployed systems can add further training objectives, multimodal components, tool interaction, retrieval, and structured control. Equally, not every chatbot is an agent. Some are simple scripted interfaces, and calling them agentic adds a claim about planning and action that they do not make.

The useful question is therefore not only "which label is correct?" It is also "what system boundary and level of abstraction are relevant to the claim being made?" A statement about model capability and a statement about end-user behaviour can both be true of the same product, and they concern different levels of description.

10. When "Ontology" Becomes a Rhetorical Weapon Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

VocabularyVocabulary can carry authority. A speaker who invokes ontology may be suggesting that their preferred classification reflects reality itself, rather than a choice made for a purpose. Before accepting that suggestion, ask:

• Which meaning of ontology is intended: metaphysical, social, or computational?
• What domain is being modelled, and what purpose does the categorisation serve?
• What entities and distinctions does the model recognise, and which does it leave out?
• Who created or maintains it, and which cases does it exclude or marginalise?
• Is it descriptive, normative, computational, or metaphysical?
• How does it handle change, uncertainty, and competing classifications?
• At what level of abstraction is the claim being made?
• What evidence would count against the proposed classification?

The point of these questions is not that every ontology is arbitrary. Every ontology involves commitments. It selects some distinctions, relations, and identity conditions as important and leaves others out, and the commitments are often most visible in what the model cannot express.

Many disputes of this kind involve one of a small set of recognisable mistakes. The usual ones are unacknowledged polysemy, where a word with several established senses is used as though it had one; shifting levels of abstraction, where a claim true of a model is applied to a system or the reverse; category confusion; equivocation, where a term changes meaning partway through an argument; undeclared ontological commitment, where a classification carries metaphysical weight that nobody has stated; and mistaking a stipulated definition for a universal essence. Naming the mistake is usually more productive than attributing a motive to the person who made it.

11. Ontologies as Situated Models FoundationalKnowledge that endures for decades — core principles

EveryEvery usable ontology is selective. A medical ontology, a legal ontology, a commercial product taxonomy, and a patient's description of an illness may all organise the same situation differently, because each serves a different purpose. Selection does not make an ontology false. A map is useful because it leaves out most of the features of the territory it describes.

Text description of the diagram

A vertical sequence of four boxes, top to bottom: World or domain; Conceptual model; Formal ontology; Data, inference, and software behaviour. Downward arrows link them, labelled in order "selection, abstraction", "formalisation", and "implementation". A dashed return arrow runs from the bottom box back up the right side to the top box, labelled "classifications reshape practice", showing that the classifications a system uses can change the world they were meant to describe.

Problems arise when the map is mistaken for the territory, when an ontology is used outside its intended domain, when its assumptions stay implicit, when historical categories are treated as timeless, when contested distinctions are presented as natural facts, or when a convenience of computation quietly becomes a claim about fundamental reality. Each of these is a failure of documentation as much as of modelling, because the assumption was never written down where a later user could see it.

A formal ontology should therefore carry, or be accompanied by, its scope, purpose, provenance, version, intended users, the competency questions it is meant to answer, its temporal validity, its known limitations, and the procedure by which it is revised. Those items are what let someone decide whether a given ontology is fit for a new use.

12. Questions to Ask When Definitions Are Disputed Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

WhenWhen two people disagree about what a term means, the following checklist tends to locate the disagreement. It can be worked through in a meeting, in a review thread, or alone.

• Which sense of the word is each person using: philosophical, social, computational, or everyday?
• What domain is being discussed, and what is the system boundary?
• What is the purpose of the classification: measurement, regulation, communication, or design?
• At what level of abstraction is each claim made: a component, a system, or a category?
• What kind of definition is on the table: lexical, stipulative, operational, technical, or persuasive?
• Does a term change meaning partway through the argument?
• What does the classification commit us to, and is that commitment stated?
• Which cases does it include, which does it exclude, and who is affected by those choices?
• What would count as evidence that the classification is wrong or unhelpful for this purpose?

13. Glossary Applied / MethodologicalKnowledge with a 5–10 year half-life — stable practice

Constructivism

The view that categories depend, at least partly, on human conceptual activity, social practice, or interpretive frameworks.

Epistemology

The study of what can be known and how knowledge is justified.

Formal ontology

An explicit, formal representation of classes, properties, relations, and constraints for a defined domain.

Interpretant

In Peirce's account, the sign or effect produced by interpreting another sign.

Language game

Wittgenstein's term for a practice within which a word's use is intelligible.

Lexical definition

A definition that records established usage.

Nominalism

The view that general categories do not exist as independent universals.

Object (Peirce)

That to which a sign refers or directs attention.

Ontological commitment

What a theory or vocabulary commits its users to believing exists.

Realism

The view that some entities, structures, or properties exist independently of anyone recognising or describing them.

Semiosis

The ongoing process by which signs produce interpretants that can themselves become signs.

Sign (Peirce)

Something that stands to somebody for something in some respect or capacity.

Stipulative definition

A meaning assigned for the purposes of a particular discussion.

Taxonomy

An arrangement of classes into broader and narrower categories, without necessarily expressing properties, constraints, or relations.

14. Conclusion FoundationalKnowledge that endures for decades — core principles

AnAn ontology does not free anyone from interpretation. It makes particular interpretations explicit and usable. Philosophical ontology asks what exists. Social ontology asks how forms of social reality exist and what a research design assumes about them. Computational ontology specifies which entities and relations a system will recognise and how it will reason with them. Semiotics explains why none of these classifications gets its meaning from a label alone: meaning emerges from signs, referents, interpretants, practices, and contexts taken together.

Definitions remain essential, especially in science, law, and engineering. A definition is not the discovery of a word's hidden essence, though. It is a disciplined proposal about how a distinction should operate within a particular setting. Productive disagreement therefore begins by moving past "words have definitions" to three questions: whose definition is it, for what purpose, and within which ontology, at what level of abstraction?

Related Topics

• Informatics, Frame Analysis, and Ontologies: Formalising "What's Going On" — the software-engineering treatment of formal ontologies and frames, which this page's sections 3 and 4 summarise rather than repeat.
• Explaining What Language Models Do: Methods and Their Limits — how the LLM category used in section 9 is explained, and where those explanations stop.
• Understanding Large Language Models — a no-code account of what the models in section 9's first structure actually do.

References

1. C. S. Peirce, Collected Papers of Charles Sanders Peirce, vol. 2, eds. C. Hartshorne and P. Weiss, Harvard University Press, 1932, §2.228.
2. L. Wittgenstein, Philosophical Investigations, trans. G. E. M. Anscombe, Blackwell, 1953, §43.
3. W. V. Quine, "On What There Is," Review of Metaphysics 2(5), 1948, pp. 21–38. Bibliographic record.
4. T. R. Gruber, "A Translation Approach to Portable Ontology Specifications," Knowledge Acquisition 5(2), 1993, pp. 199–220. https://doi.org/10.1006/knac.1993.1008
5. N. Guarino (ed.), Formal Ontology in Information Systems: Proceedings of the First International Conference (FOIS'98), Frontiers in Artificial Intelligence and Applications vol. 46, IOS Press, 1998.
6. J. R. Searle, The Construction of Social Reality, Free Press, 1995.
7. I. Hacking, The Social Construction of What?, Harvard University Press, 1999.
8. W3C, OWL 2 Web Ontology Language Document Overview (Second Edition), W3C Recommendation, 11 December 2012. https://www.w3.org/TR/owl2-overview/
9. A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, I. Polosukhin, "Attention Is All You Need," Advances in Neural Information Processing Systems 30 (NeurIPS 2017). https://arxiv.org/abs/1706.03762