Observer and Model: How Much Does Our Description Shape What We See?
Observation, measurement, and modeling are not assumption-free windows. This article shows how variable selection, operationalization, scale, and measurement procedures shape description without implying that reality is arbitrarily created by the observer.
When we observe the world, we do not receive it as a perfectly unprocessed stream of facts. We choose what to measure, with which instrument, at what scale, under which conditions, and with which model the result will be interpreted. Even a simple graph is therefore not the world itself; it is a selected representation of some features of the world.
This does not mean that reality is invented at will. A good model may simplify, idealize, and omit, but for its intended purpose it must remain in contact with something that can surprise, constrain, or falsify it. A meter, equation, or classification does not create a planet, disease, or temperature merely by describing it in a particular way.
The better question is more precise: how much of a result comes from the world, how much from the way it is measured, and how much from the conceptual framework used to organize it? Science does not solve this problem by removing the observer. It addresses it by making procedures increasingly explicit, testable, and reproducible.
The observer and the model are therefore parts of the route to knowledge, not necessarily obstacles to it. The danger begins when we forget which properties we selected, which assumptions we built in, and where a useful map starts presenting itself as the whole territory.
A description is not the thing described
Scientific descriptions use words, numbers, graphs, equations, images, simulations, and models. None of these representations is a literal copy of its target. A subway map deliberately distorts distances and angles because its job is to display connections between stations clearly. In the same way, a scientific model may ignore friction, treat a population through averages, or select only a few variables because it is built to answer a particular question.
The question “is the model true?” therefore often needs a qualifier: true or adequate in what sense, and for which purpose? A free-fall model can be highly useful over short distances even while neglecting air resistance. A climate model does not need the position of every molecule in order to be useful at a different level of description.
Purpose, however, does not justify everything. If a model predicts the wrong outcomes, fails on data outside those used to build it, or depends on assumptions that conflict with well-supported measurements, its usefulness declines. The choice of representation is human; the world's response to our predictions and interventions is not merely a matter of choice.
Measurement is more than reading a number
Measurement often looks direct: an instrument displays a number and we record it. In practice, several steps lie between a phenomenon and a result. We must specify the quantity of interest, select a procedure, calibrate the system, estimate environmental influences, and transform an instrument signal into a value attributed to the property being measured.
The International Vocabulary of Metrology therefore treats measurement as an experimental process for obtaining values that can reasonably be attributed to a quantity, not as simply looking at a pointer. A measurement result normally also carries relevant information about uncertainty. In demanding measurements, the relation between instrument signal and target quantity explicitly depends on a measurement model.
This does not weaken measurement. It is precisely what makes improvement possible. Calibration can be repeated, one model can be compared with another, uncertainty can be evaluated, conditions can be standardized, and a result can be checked in another laboratory. Objectivity is not the absence of procedure; it is often achieved by making the procedure public, traceable, and testable.
Operationalization: how a concept becomes a variable
For some quantities, the path from concept to measurement is comparatively direct. For others, an abstract idea must be translated into observable indicators. This is operationalization. “Poverty,” “trust,” “intelligence,” “political stability,” or even “quality of life” do not come with a single natural indicator waiting to be read off from the world.
A chosen indicator may capture an important part of a construct while omitting another. An income threshold does not capture every dimension of material security; a questionnaire may measure part of a psychological trait, but its meaning depends on the validity of the interpretation placed on the scores. It is therefore not enough to say that something is “measured.” We must ask exactly what was operationalized and what evidence supports the claim that the indicator carries the meaning assigned to it.
The history of operationalism also reveals the limit of the strongest version of this idea. A concept need not be identified with one and only one measurement operation. Different operations can approach the same construct from different directions, and their agreement or disagreement can itself become part of the research question.
Choosing variables and scale determines what a model can see
Every model highlights some features and pushes others into the background. If an epidemiological analysis omits age, an important pattern may disappear. If an economic model collapses very different households into a single average, it may lose the distribution of effects. If weather is examined at hourly, daily, or decadal scales, the resulting descriptions can differ without necessarily contradicting one another.
Scale is therefore a substantive choice. A molecular description of a gas and a thermodynamic description of pressure do not necessarily compete to be the one correct picture. They operate at different levels and answer different questions. Likewise, a road map and a geological map of the same region are not relativistic rivals; each preserves a different set of relations.
The problem begins when a model's boundary is mistaken for reality's boundary. If a phenomenon is absent from a dataset, it does not follow that it is absent from the world. It may be unmeasured, misclassified, or lost through coarse resolution. Good analysis therefore states not only what the model contains, but also what it leaves out.
Observation can be theory-laden without being arbitrary
A measuring instrument does not interpret its own signal. A thermometer is informative because we have a theory of its response, calibration, and relation between the reading and temperature. A spectrum from a distant star becomes evidence about composition only through physical knowledge of light, atoms, the instrument, and data processing.
Philosophers of science often describe this as the theory-ladenness of observation. Empirical results are not completely separable from the concepts, theories, and models used to produce and interpret them. But it does not follow that a measurement can mean anything whatever. Bad calibration, failed prediction, and unsuccessful replication remain problems even after the role of theory is acknowledged.
Science reduces the danger of circularity through independent measurements, different instruments, blinded procedures, standards, replications, and tests in which a model risks failure. Theory helps data become evidence; data can in turn force the theory to change.
The observer effect: when measuring changes what is measured
The expression “observer effect” has a narrower scientific meaning than it often receives in popular discussion. In some cases, the act of measurement changes the system. A thermometer exchanges heat with an object, a probe can alter a flow, microscope illumination can damage a sensitive sample, and asking a survey question can influence a later response.
This is not the same claim as saying that an observer's consciousness creates external reality. The effect can be entirely physical or behavioral: the measuring device, experimental intervention, or awareness of being observed changes the process under study. A central task of experimental design is to estimate the size of this influence and determine whether it can be reduced, corrected, or modeled.
The observer effect should also be distinguished from observer bias. In the former, the system itself may change; in the latter, the observer's expectations affect recording, selection, or interpretation. The remedies differ: one may require a better instrument or protocol, while the other may be reduced by blinding, predefined coding rules, and independent checking.
The quantum “observer” is not a shortcut to metaphysics
The word “observation” becomes especially hazardous when moved into quantum mechanics without explanation. Quantum measurement is indeed a deep conceptual problem, but the formalism ordinarily describes interactions between a system and measuring apparatus and the production of measurable outcomes. Experiments such as the double slit show that obtaining which-path information can alter an interference pattern.
That does not establish a general scientific conclusion that human consciousness creates matter or that thought arbitrarily selects physical reality. The history of quantum interpretation includes proposals that assigned consciousness a special role, but those proposals are not the same thing as an experimentally established consequence of quantum mechanics.
This is why disciplined language matters: measurement context, physical interaction, information, decoherence, and philosophical interpretation are not synonyms. The detailed quantum measurement problem deserves its own treatment; for the present argument the important boundary is that the word “observer” does not by itself complete a metaphysical argument.
Model dependence is not the same as relativism
Once we acknowledge that different models select different variables and idealizations, it is tempting to conclude that truth is merely a matter of perspective. That conclusion is too quick. Multiple representations of the same system can be legitimate because they serve different tasks without all representations becoming equally good.
A model may have a limited domain of validity and still be objectively better than a rival at predicting a specified class of phenomena. One model may be more precise, another more transparent, and a third more useful for interventions. Sometimes models complement one another; sometimes they genuinely conflict and further measurements or experiments must discriminate between them.
We can therefore accept the perspectival character of knowledge without embracing arbitrary relativism. We may not possess a literal “view from nowhere,” but we do have ways to compare perspectives: empirical adequacy, predictive success, robustness, agreement with independent measurements, explanatory power, and clearly stated limits of application.
How to test how much of a result comes from the model
The first question is the model's goal. Prediction, explanation, classification, measurement, and decision support are not the same task. Next ask which variables were selected, which were omitted, and why. Would another reasonable operationalization produce a similar result?
The second question is sensitivity. If a small change in assumptions reverses the conclusion, the result is more model-dependent than one that survives several reasonable specifications. Robustness is not proof of truth, but it helps reveal whether a conclusion hangs on one arbitrary choice.
The third question is contact with independent data. Does the model predict something it did not use while being built? Does the result replicate with another instrument, in another population, or by another method? Can the model state where it expects to fail? A model that only explains data already seen is epistemically weaker than one that successfully risks new predictions.
The most defensible picture: world, measurement, and model interact
The most useful picture is neither naïve realism, in which data speak entirely for themselves, nor radical relativism, in which our description determines everything. An empirical result emerges from a relation among phenomenon, measuring system, procedure, model, and research question.
The world contributes constraints. Instruments and methods determine which traces are detected. Models determine how those traces are organized. Observers choose questions and must take responsibility for assumptions. Then comes the crucial step: the result is exposed to other measurements, other models, and the possibility that it is wrong.
A good model is therefore not one that pretends to have no perspective. A good model knows its perspective, states its purpose, exposes its assumptions, and allows reality to push back. Description really does shape what we are able to see — but that does not make description the master of everything that exists.
Sources and further reading
- THY-REALITY — Ena resničnost, mnogo zemljevidov / One Reality, Many Maps (LOCKED): representation, purpose, idealization and model pluralism without relativism.
- THY-REALITY — Dejstvo, interpretacija, hipoteza in špekulacija niso isto / Fact, Interpretation, Hypothesis and Speculation Are Not the Same (LOCKED): separation of observation from interpretive claims.
- THY-REALITY — Ko družba postane vmesnik: merjenje kolektivne resničnosti / When Society Becomes an Interface: Measuring Collective Reality (LOCKED): indicators, categories, measurement feedback and metric effects.
- THY-REALITY — Kaj sploh pomeni resničnost? / What Do We Mean by Reality? (LOCKED): world, experience, model and description as distinct levels.
- THY-REALITY — Percepcija ni resničnost: kako možgani gradijo naš svet / Perception Is Not Reality: How the Brain Builds Our World (LOCKED): construction of perception without denying an external world.
- THY-REALITY — Kje se konča dejstvo in začne interpretacija? / Where Does Fact End and Interpretation Begin? (LOCKED): measurement, provenance and inference are method-dependent but auditable.
- Frigg, R.; Nguyen, J. — Scientific Representation. Stanford Encyclopedia of Philosophy, substantive revision 2026. Scientific representations and models need not be literal copies of their targets; representation is purpose- and practice-sensitive.
- Bogen, J. — Theory and Observation in Science. Stanford Encyclopedia of Philosophy, substantive revision 2026. Theory-ladenness of observations, instrument-mediated data and the epistemic role of empirical evidence.
- Tal, E. — Measurement in Science. Stanford Encyclopedia of Philosophy. Measurement as model-, theory- and procedure-dependent while remaining capable of estimating mind-independent quantities.
- Chang, H. — Operationalism. Stanford Encyclopedia of Philosophy. Bridgman, operational definitions and the limits of identifying a concept with a single measurement operation.
- JCGM / BIPM — International Vocabulary of Metrology (VIM3), entries 2.1, 2.9 and 2.26. Measurement, measurement result and measurement uncertainty.
- Possolo, A. — Simple Guide for Evaluating and Expressing the Uncertainty of NIST Measurement Results. NIST Technical Note 1900 (2015). Measurement models, inputs and probabilistic uncertainty.
- Cronbach, L. J.; Meehl, P. E. — Construct Validity in Psychological Tests. Psychological Bulletin 52 (1955), 281–302. Construct validation as an evidential network rather than identification with a single operational measure.
- Massimi, M. — Perspectival Modeling. Philosophy of Science 85(3), 335–359 (2018). Plural and apparently incompatible models can play exploratory roles without requiring arbitrary relativism.
- Kincaid, J.; McLelland, K.; Zwolak, M. — Measurement-induced decoherence and information in double-slit interference. American Journal of Physics 84(7) (2016), via NIST. Measurement interaction, which-path information and loss of interference.
- Stanford Encyclopedia of Philosophy — Philosophical Issues in Quantum Theory. The quantum measurement problem concerns system-apparatus dynamics and interpretation; the term observer does not by itself establish a consciousness-causes-reality thesis.