Scientific models are neither facts nor mere theories: they are calibrated representations that can be factually reliable in some domains and speculative in others. A model counts as factual when its predictions match measurement within stated error bars, and as theoretical when its assumptions outrun evidence. The distinction is not whether something is called a theory, but whether it can be tested and at what scale it holds. A climate model correctly predicting 1.5 F of warming is not the same as a cosmology simulation that is unfalsifiable in practice.
What a Scientific Model Actually Is
A scientific model is a simplification built to make predictions. The model is not the thing. A weather model is not the atmosphere; it is a set of equations that approximate the atmosphere at a chosen resolution. The US National Weather Service's Global Forecast System runs at roughly 13 km (8 mi) horizontal grid spacing. Anything smaller than a grid cell is averaged, parameterized, or ignored. That averaging is the model's core method and its core limitation.
Models sit on a ladder of abstraction. A physical model like a wind tunnel is the most direct. A mathematical model like Newton's laws sits one level up. A statistical model like a regression fits numbers without claiming causation. A conceptual model like a food web diagram is the least quantified. Each rung loses something and buys something. In plain terms: a model is a map chosen for a purpose, and the purpose sets the scale.
The crucial mechanism is parameterization. Where equations cannot resolve a process at grid scale, modelers substitute a formula fitted to observations. Cloud microphysics in a climate model is parameterized. The model does not simulate each droplet; it approximates a droplet population's effect on radiation and rain. This is why two models can be equally defensible and disagree by a factor of two in regional rainfall.
Facts, Theories, and Models: The Vocabulary Problem
The word fact in science means a repeatable observation: water boils at 212 F at sea level. The word theory means a well-tested explanatory framework: the theory of evolution, the theory of relativity. A model is a working instrument that sits between them. Confusing these three terms is the source of most public argument about whether models count as facts.
Calling a model a theory misses the point. A model can be validated without the underlying framework being confirmed, and a confirmed framework can produce poor local models. General relativity is well supported, but a regional gravity model can still be wrong because of a bad terrain dataset. The framework and the local instrument are separate claims.
A professional caveat matters here. Model validation is domain-specific. A model validated at 10 km resolution may fail at 1 km. The model does not become false; its range of validity narrows. The honest statement is "accurate within this range under these assumptions," never a bare "proven."
The Comparison: Model Types and Their Evidentiary Status
Different model types carry different evidentiary weight because they are tested differently. The table below compares six common types by what observation actually supports them.
| Model type | Typical test | Evidentiary status | Example error range |
|---|---|---|---|
| Physical scale model | Direct measurement | Factual within calibrated range | 5-15 percent force error |
| Deterministic equation model | Prediction vs. measurement | Strong if validated | Moon position, meters over decades |
| Statistical fit model | Held-out data test | Correlational only | 60-90 percent variance explained |
| Numerical climate model | Hindcast plus multi-model ensemble | Probabilistic, not exact | Global mean 1-2 F spread |
| Cosmological simulation | Consistency with survey data | Hypothesis-level in detail | Not directly falsifiable locally |
| Conceptual diagram model | Qualitative fit | Heuristic, not predictive | No numeric error bar |
The key quantified comparison is the last two columns. A deterministic model like orbital mechanics delivers accuracy measured in meters over decades. A climate model delivers a range, not a point, and its value lies in that range being stable across independent models. The comparison shows a model's status is a continuum, not a binary.
A Timeline of How Model Status Changes
Model status is not fixed. New data and new compute move models up or down the ladder. The timeline shows representative shifts.
| Period | Development | Impact on model status |
|---|---|---|
| 1687 | Newton's laws published | Deterministic models become testable |
| 1900s | Statistical regression formalized | Fit models gain predictive standing |
| 1950s | First numerical weather forecast | Computation enters prediction |
| 1960s | General circulation climate models | Ensemble method emerges |
| 1990s | Satellite data assimilation | Hindcast validation grows |
| 2010s | Multi-model ensembles standardized | Uncertainty ranges become the product |
| 2020s | Machine-learned emulators | New class with opaque assumptions |
Each row shows a method change that altered what counted as evidence. The 1950s shift matters most for the public question. Before numerical forecasting, weather prediction was mostly pattern-based. After it, the model's output became a falsifiable claim. That change is why weather models are now held to factual standards while many economic models are not.
Why Models Get Called Facts and Theories at the Same Time
The same model can be factual in one role and theoretical in another. A climate model's historical forecast can be scored against 100 years of records. Its 2100 projection cannot be scored yet. The model is validated and unvalidated depending on which output you read. People cite whichever verdict supports their position.
Another non-obvious mechanism is ensemble logic. A single model output is a theory about one possibility. The mean of 30 independent models is a statistical estimate with a known spread. The spread is the honest output. When a model range is 2-5 F of warming, the range is the finding, not a weakness. Reporting the midpoint alone as a fact misuses the model.
A corrected misconception: a model being wrong does not disprove the framework behind it. The 2021 ocean-heat model errors did not overturn thermodynamics. They exposed a bad mixing parameter and were corrected in the next generation. Models fail at the instrument level far more often than at the law level. Distinguishing the two failures is the core skill.
Where Models Break: Edge Cases and Limits
Models break predictably at boundary conditions. A model validated in the tropics often fails near the poles because the underlying parameterizations were tuned on tropical data. A financial risk model calibrated on a decade of calm markets fails in a once-in-fifty-year event. The failure is not random; it concentrates where the training data are thin.
Models also break when they are used outside their scale. A 1 km flood model run at 100 m resolution will produce artifacts, not more accuracy. Resolution increase costs compute quadratically while accuracy improves linearly at best. This is why climate model centers run 25 km to 100 km grids, not 1 km, for long projections.
Red flags that a model claim is overstated:
- A single run presented as a prediction without an uncertainty range.
- The model's validation dataset overlaps its calibration dataset.
- The claim extends a model from its training domain into a new region or regime.
- Parameter values are cited without their source or calibration period.
- A 99 percent confidence claim on a model with a 10 percent historical error.
Any one of these should reduce trust in the model's factual status.
What to Actually Do When You Read a Model Claim
Ask five concrete questions. First, what is the model's validation period? A model validated over 20 years is more credible than one validated over two. Second, what are the error bars? A number without a range is a guess, not a result. Third, what was held out? If everything was used for calibration, the model was never truly tested.
Fourth, what is the resolution or scale? Match the claim to the grid size. A 25 km model cannot make a neighborhood-level claim. Fifth, was it an ensemble? A single run is one draw, not a finding. These five checks turn a binary fact-versus-theory argument into a graded judgment.
Apply the same checklist to models you rely on indirectly. A pollution model behind a regulation, a credit model behind a loan decision, and a medical risk score all invite the same questions. The model's authority comes from its documented test, not its complexity. A simple model with a held-out test beats a complex model with none.
Why the Fact-or-Theory Framing Fails
The question assumes a binary the evidence does not support. Models occupy a graded position defined by validation, scale, and purpose. A well-tested model is a factual instrument within its range. An untested model is a hypothesis regardless of its mathematical sophistication. The right question is not "fact or theory" but "validated where, how well, and against what."
This reframing has practical consequences for public debate and for policy. It shifts attention from labels to test results. It makes uncertainty ranges a feature to be read, not a flaw to be attacked. A reader who asks "what was held out" is harder to mislead than one who asks "is it a fact." The mechanism-level answer is the durable one: models are instruments, and instruments are judged by calibration.