Maps, Models, and Understanding
All empirical knowledge is model-mediated
The London Underground map is, by any geographic standard, a tissue of falsehoods. Distances are wrong. Directions are wrong. Lines that curve through the city run straight on the diagram; stations miles apart sit next to each other; the Thames is a stylized ribbon. Everything a surveyor would demand, the map discards — and it is one of the most successful maps ever drawn, precisely because of what it discards. A passenger needs to know which line reaches which station and where to change trains. The Tube map preserves exactly those structural relations and deliberately distorts everything else. The distortion is not a defect of the design. The distortion is the design.
Every model works this way. A model is defined by two choices: what to preserve and what to throw away. Newtonian mechanics throws away relativistic corrections and quantum effects, and captures the dynamics of nearly everything you will ever push, drop, or drive. General relativity throws away Newton’s forces for a geometric account that keeps working where Newton’s fails. Neither is a photograph of reality; each is a Tube map of it, preserving the invariants that matter for a domain and simplifying the rest. So the right question about a model is never is it true, full stop? but which structural features does it preserve, over which domain, at what cost?
Alfred Korzybski compressed the caution into a slogan: the map is not the territory. As a cognitive safeguard the slogan is invaluable, and I will come back to the catastrophes that follow from forgetting it. But on its own it undersells the situation, in two directions at once. It undersells the maps: “not the territory” makes distortion sound like a regrettable gap when, as the Tube map shows, it is the whole source of a model’s power — adequacy is domain-specific, fixed by what the model chooses to preserve. And it undersells the predicament: the slogan invites the fantasy that if we were careful enough we could put the maps down and inspect the territory directly. We cannot. There is no unmediated access to reality. All empirical knowledge — scientific and ordinary alike — comes to us through models, and it could not be otherwise.
No View Without a Map
The dream of foundational, irrefutable knowledge — certainty secured before inquiry begins — does not describe anything cognition actually does. From infancy onward, minds run on structured expectations. An infant forms rudimentary causal expectations about objects and motion; a child revises internal representations through interaction; an adult reads social behavior through predictive heuristics that no one ever taught explicitly. In every case the agent deals not with an unfiltered external reality but with a structured representation of it — a model that organizes sensory input, imposes explanatory structure, and supports prediction. Perception itself is already interpretation.
Scientific modeling does not escape this architecture; it formalizes it. When cosmologists analyze the cosmic microwave background, they do not read off facts about the early universe the way one reads a thermometer. They compute what a parameterized model of early-universe physics predicts the data should look like, compare, and adjust. The parameters they report are creatures of the model. A statement such as “the expansion rate of the universe is \(H\)” is shorthand for a conditional: if the data, the modeling assumptions, and the parameterization of cosmology are accepted, then the estimate for \(H\) takes a specific value. There is no model-independent interpretation of the observations on offer — not because cosmologists are sloppy, but because “interpretation without a model” is not a coherent request.
This is Conditionalism’s home territory. If all truth is conditional — if truth values attach only to statements evaluated given their background conditions — then model-mediation is not an unfortunate limitation on empirical knowledge but its normal form. The model is where the conditions live. And since we can never hold a statement up against unmediated reality to check the match directly, empirical truth is assessed as coherence within a model that successfully organizes and predicts experience — the working level of the three levels of truth. Understanding, on this account, is not a mystical rapport with things in themselves. Understanding is the construction and refinement of models.
How Models Earn Their Keep
If models are neither deduced from self-evident axioms nor read off from raw experience, where do they come from, and why should we trust any of them?
They come from three kinds of inference working in a loop. Abduction generates candidates: given puzzling observations, infer the structure that would render them intelligible. Deduction develops consequences: given a candidate model, work out what else it commits us to. Induction keeps score: check the developed consequences against observation and redistribute confidence among the candidates accordingly. Deduction is the only one of the three that yields necessity, and it is also the only one that adds no new content — it explores assumptions already in place. The creative and evaluative work is done by the other two.
That evaluative work is Bayesian in structure, whether or not the scientist writes down a prior. Scientific inference is the updating of Credences — degrees of belief — across competing models. The inverse-square law of gravity was never proven; it earned overwhelming Credence because it rendered the observed motions highly probable while its rivals rendered them improbable. General relativity displaced Newtonian gravity by the same tribunal: Mercury’s perihelion shift and the bending of starlight were exactly what Einstein’s model predicted and Newton’s could not accommodate. I defend this framework at length in defense of Bayes; here I need only the structural point. A reported result like “the Hubble constant is \(67 \pm 3\) km/s/Mpc” summarizes a posterior distribution conditioned on a cosmological model, a body of data, and assumptions about measurement error. The frequentist alternative — error bars characterizing the long-run behavior of hypothetical repeated experiments — can be mathematically impeccable while answering a question nobody asked. The Bayesian statement answers the question researchers actually pose: given this model and these data, what parameter values should we take seriously?
Probability, in this role, is the quantitative expression of conditional truth. It measures how an agent distributes Credence across models — a fact about the agent’s epistemic state, not about any objective weight in the world. That objective weight exists too: the Measure of branches in the Quantum Branching Universe (QBU), the Everettian picture I develop in Measure and Credence. Keeping the two separate is the root discipline of probabilistic reasoning, and model evaluation lives entirely on the Credence side of the ledger.
The model-dependence goes deeper than parameter estimates; it reaches into physical law. A gas contains an astronomical number of degrees of freedom whose microstate evolves deterministically but is computationally inaccessible, so we describe it instead with coarse-grained macrostates — temperature, pressure, density — each standing for a probability distribution over microstates. Entropy measures the limitations of that description: how much work can be extracted given the information available at the chosen level of coarse-graining. The second law of thermodynamics is best understood not as a brute fact about matter but as a constraint arising from the level of description we have chosen — a law about our maps that is nonetheless as reliable as anything in physics. Quantum mechanics presses the same lesson from the other side: the interpretations that compete to make sense of the formalism — Copenhagen, Everett, QBism — agree on every operational prediction and differ as meta-models, alternative choices for connecting the same probabilistic structure to an ontology. Choosing among them is itself an exercise in model evaluation, and I make my choice in probability without collapse.
Even the frontier disputes of cosmology are disputes between maps. The Hubble tension — CMB-based inferences of the expansion rate disagreeing with local distance-ladder measurements — is not reality contradicting itself. It is two model stacks, each embedding assumptions about astrophysics, instrument calibration, and priors, delivering incompatible posteriors. Resolving it means finding which modeling assumption is inadequate. That is what scientific crisis normally looks like from the inside: not the territory misbehaving, but a map reaching the edge of its domain.
When the Map Is Mistaken for the Territory
Compression is foundational to cognition — an incompressed representation of reality would be as useless as a 1:1 map — but the same necessity breeds a family of characteristic errors. This is where Korzybski’s slogan does its work, as a pathology catalogue.
Labeling as explanation. Naming a phenomenon — “depression,” “inflation,” “intelligence” — produces a feeling of understanding without any predictive gain. A label is a placeholder on the map, not a mechanism in the territory. The test is always the same: does the term let you predict anything you could not predict before?
Reifying abstractions. Economic indicators, personality scores, and political categories are summary statistics of underlying processes, yet they are routinely treated as the underlying things themselves — the abstraction promoted to foundation, and then defended as if reality owed it allegiance.
Elevating models to laws. Newtonian mechanics succeeded so spectacularly that its domain-specific adequacy was mistaken for universal necessity — until Einstein and the quantum showed where the map ran out. Every model has breaking points; the error is not in trusting a model but in forgetting that its warranty is conditional on its domain.
Entrenched dogmatism. Ideological maps routinely outlive their predictive validity, kept alive by confirmation bias, cognitive inertia, and social conformity. A map maintained for reasons other than its fit to the territory has stopped being a map and become a badge.
Pretending completeness. A perfect map is impossible in principle — completeness would mean replicating the territory, which defeats the purpose of having a map. Every model omits; the only question is whether you know what yours omits. The dangerous cartographer is not the one whose map has blank regions but the one whose map has none.
Each failure has the same anatomy: a conditional truth stripped of its conditions and worshipped as an absolute. The cure is correspondingly uniform — drag the conditions back into the light and ask what the model preserves, what it distorts, and where its domain ends.
Better Cartography
What makes one map better than another? Not comfort, not elegance, not ideological fit: predictive success and practical utility within a stated domain. Models are pragmatic instruments, and the standards for judging them are the pragmatic and coherence levels of truth doing their ordinary jobs — does it predict, does it integrate, does it survive contact with the anomalies?
Mastery, though, is not possession of the one best map; it is fluency across several. Physicists move between particle and wave descriptions as the problem demands. Economists triangulate among frameworks none of which is adequate alone. The skill of selecting a map — knowing which representation fits which purpose — is itself an intellectual discipline, and treating map-choice as betrayal of the One True Map is how dogmatists are made. Cartographic pluralism is not relativism: the maps are ranked, per domain, by performance, and the ranking is as objective as evidence can make it.
The maintenance regime is a standing interrogation of your own frameworks. Is this label explanatory, or merely descriptive? Is this model predictive, or just familiar? What contradictory evidence am I discounting, and what is my map omitting? Above all: am I prepared to abandon it? If the honest answer is no, the collision with reality is not avoided, only scheduled — the habit of revising in time is the discipline of updating.
None of this makes the territory less real. Model-mediation is an epistemic thesis, not an idealist one: gravity operates independently of anyone’s belief in physics, and economies collapse regardless of the ideological commitments of their managers. The territory is exactly what pushes back when a map fails — which is why the failure of maps is evidence for the territory, an argument I complete in conditional realism. The map is where “the world” becomes intelligible to us; the world itself remains mercilessly indifferent to our representations of it.
That pairing — no access except through models, no mercy for bad ones — is the whole case for cognitive humility, and for its active form. Understanding is not the end-state of finally seeing reality bare; there is no such state. Understanding is the ongoing construction of models that predict, integrate, and explain better than their predecessors, held with exactly the confidence their performance has earned and not a degree more. The disciplined mind does not seek maps that feel like the territory. It seeks maps that work — refined, validated, contextually appropriate — and keeps the revision pencil sharp.