The History of Similarity
Similarity is so familiar that it barely registers as a concept. We casually judge objects, ideas, and situations to be similar, and we act on those judgments with confidence. In modern technical systems — especially AI-powered search, retrieval, and recommendation — similarity quietly serves as an intellectual foundation.
Yet the apparent simplicity of similarity is deceptive. Similarity has a long and consequential intellectual history. For most of Western thought, similarity was treated as a fact about the world itself: something discovered, not established. Only recently has similarity come to be understood as dependent on context, purpose, and use.
This shift did not arrive suddenly with artificial intelligence or modern computing. It unfolded slowly, tracking deeper changes in how philosophers understood knowledge, meaning, and action. What follows is a brief tour of that evolution.
Forms and Universals
For the ancient Greeks, similarity reflected what things truly were.
Plato grounded similarity in Forms and Universals. Objects were similar insofar as they imperfectly instantiated the same abstract ideal. Similarity, in this view, was metaphysical rather than practical. Context, purpose, and use were irrelevant.
Aristotle rejected transcendent Forms but preserved the assumption that similarity was intrinsic. His theory of Universals located similarity in shared essential properties that defined natural kinds. Similarity enabled classification and taxonomy, precisely because it was anchored in the structure of reality itself.
For nearly two millennia, this essentialist framing dominated Western thought. Similarity was something to be uncovered, not constructed.
Mental Constructs
The first serious cracks appeared in the Middle Ages, when philosophers began to question whether universals existed independently of the mind.
William of Ockham argued that universals were mental constructs rather than features of the external world. On this view, similarity reflected patterns imposed by human cognition rather than facts embedded in reality. Similarity began to move from the world into the mind.
Early modern philosophy pushed this further. John Locke distinguished between primary qualities, which belonged to objects themselves, and secondary qualities, which depended on perception. Similarity judgments increasingly appeared to depend on which qualities observers attended to, and for what purpose.
This was a profound shift, but it remained incomplete. Even as similarity became psychologically mediated, philosophers largely assumed that a true, objective similarity existed beneath subjective perceptual distortion.
Function over Form
In the nineteenth century, similarity began to be framed in terms of function rather than structure.
Charles Darwin made this unavoidable. In evolutionary theory, traits could be similar because they served the same function, even if they arose independently. The wings of bats and birds were similar for flight, not because they shared an essence or a common origin.
Here, similarity depended on explanatory context. Were we asking about ancestry, adaptation, or survival? The answer determined what counted as similar. Similarity became instrumental rather than purely descriptive. This reframing weakened the idea that similarity was a single, privileged relation waiting to be discovered.
Contextual Similarity
The twentieth century completed the transition from intrinsic similarity to contextual similarity.
Ludwig Wittgenstein rejected the idea that categories required shared defining features. His notion of family resemblance showed how similarity could emerge from overlapping uses, with no single common property. Similarity was grounded in practice, not essence.
Cognitive science added empirical support. Eleanor Rosch demonstrated that people organize categories around prototypes rather than strict boundaries. Similarity is graded, not binary, varying with task and context.
Amos Tversky went further, showing that similarity judgments depend on weighted features whose importance shifts with goals. Similarity can even be asymmetric, violating classical assumptions altogether.
The ancient picture was inverted. Similarity was no longer grounded in universals. It was something people established in the course of acting in the world.
A Persistent Tension
The history of similarity is not a story of replacement, but of accumulation.
Intrinsic similarity remains indispensable for representation, compression, and generalization. Contextual similarity is essential for relevance, decision-making, and utility. Problems arise when one is asked to do the work of the other.
This tension persists because similarity is pulled in two directions: toward truth and toward use. The modern insight is not merely that similarity is contextual, but that it always answers a question — whether or not that question is made explicit. Every operational notion of similarity implicitly encodes a theory of what matters.
Conclusion
Similarity began as an assumption about the structure of reality. Over centuries, it became a reflection of perception, then function, and finally use. The modern conception of similarity — especially in artificial intelligence and related technical systems — is historically young, but it addresses problems earlier frameworks could not even articulate.
To understand similarity, then, is not simply to compare objects. It is to ask what the comparison is for.
p.s. A contemporary echo of the Platonic intuition appears in the Platonic Representation Hypothesis (PRH). Modern deep networks, trained independently across tasks and modalities, develop aligned high-dimensional embeddings in which similarity is encoded geometrically. This suggests the existence of a shared latent structure — an “ideal” representation of the world — that models approximate. In effect, PRH operationalizes the ancient Platonic idea of universal forms, but in an empirical, measurable form: similarity is not categorical or metaphysical, but emergent, graded, and testable. Seen this way, PRH represents the latest point in the long trajectory from Plato’s ideal Forms, through Aristotle’s properties, to modern cognitive and computational conceptions of similarity. Thanks to Russell Jurney for pointing me in this direction!
