By G. Sawatzky, knowledge-foundation.ai
Originally published: August 27, 2025
Recent AI systems, especially large language models (LLMs), have renewed interest in an older question about knowledge representation: how much of meaning depends on strict categorization, and how much depends on overlapping patterns of use and similarity? These systems often connect concepts successfully without relying on a rigid taxonomic structure. That makes it worth revisiting philosophical approaches that challenged the idea that every concept must be defined by a perfect set of necessary and sufficient conditions.
For much of the computer age, knowledge representation has been shaped by a strong preference for precision. Systems were designed with exact database schemas and knowledge bases that required everything to fit into predefined categories. That approach has produced many valuable tools, but it also creates real difficulties. Concepts in the world often resist neat categorization, and boundary cases quickly expose the limits of rigid taxonomies.
Classic examples highlight this issue: Is a tomato a fruit or a vegetable? Is a PhD student an employee, a student, or both? These questions expose the limitations of a purely rigid, taxonomic approach. The quest for perfect knowledge models becomes a struggle against the messy and interconnected nature of reality.
Philosophers have long explored this problem. The Austrian philosopher Ludwig Wittgenstein introduced the concept of "family resemblances" to describe how things in the world are related. In his view, a category is not defined by a single set of shared characteristics but by a series of overlapping similarities. Think of the members of a family: while no single trait is shared by all of them, they are connected by a web of common features like a similar nose shape, eye color, or speaking style.
That point should not be conflated too quickly with later technical work. Wittgenstein's argument is philosophical: concepts do not always hold together because they share one essence. Computer scientists such as Joseph Goguen and John Sowa are relevant here for a different reason. They explored formalisms and representational strategies that better accommodate inexact concepts, context, and structural complexity than a strictly rigid classification scheme.
Goguen, through his work on inexact concepts and related ideas, emphasized that categories can be understood in terms richer than a simple true-or-false classification. Sowa, a major figure in knowledge representation, criticized the limits of purely formal logical systems when they are treated as if they exhaust how meaning works in practice. So the connection is not that Goguen simply "applies" Wittgenstein, but that both help expose the limits of overly rigid categorization from different directions.
Large language models have, in practice, made this issue harder to ignore. They do not operate through explicit, rule-based logic in the traditional sense. Instead, they process large amounts of data and learn patterns of similarity across contexts of use. When an LLM treats a "student information system" and an "educational management system" as conceptually related, it is usually not because both terms were tied to one exact rule. It is because the surrounding language shows a meaningful overlap in use.
This ability allows AI to act as a bridge between disparate systems that lack perfect, shared definitions. For example, a healthcare system that calls a record a "patient identifier" and a second system that calls it a "member ID" might be difficult to link through traditional, rule-based logic alone. An AI, by recognizing the family resemblance between the two terms based on their context and use, can sometimes establish a meaningful connection.
This suggests a different view of interoperability. Instead of assuming that precision always requires perfect categorization, we can sometimes achieve useful alignment by recognizing stable patterns of resemblance and use. That does not remove the need for rigor, but it does shift where some of that rigor has to live.
The argument for family resemblances does not mean we should abandon foundational data-modeling principles. A well-defined conceptual model is still essential for building reliable systems. A methodology like Object Role Modeling (ORM) remains valuable precisely because it captures business rules and relationships in an explicit, human-readable form. For related arguments about conceptual schema and ontology, see What Does an Ontology Actually Give You? and An ORM-Based Semantic Framework.
The concepts discussed in this article apply primarily to external interoperability. When two systems, each with its own precise internal model, need to communicate, their schemas often do not align perfectly. Instead of forcing a brittle one-to-one mapping, we can sometimes use AI's ability to recognize family resemblances across terms, roles, and contexts. This allows a more flexible connection while preserving the internal integrity of each system.
One practical direction is to share conceptual models in lightweight representations such as JSON or YAML and let AI systems reason over their overlapping structures and verbalizations. That possibility also sits naturally beside the broader distinction between information, semantics, and knowledge discussed in What I Mean by Knowledge, Information, and Semantics.
The point is not that "good enough" replaces rigor. It is that rigor may need to be distributed differently: exact inside a local model, and resemblance-based at the boundary between models. That shift may be one of the more useful lessons AI is forcing us to reconsider.
© 2025 G. Sawatzky. Licensed under CC BY-NC-ND 4.0.