Home / Philosophy of Science | What Makes Knowledge Scientific?

Philosophy of Science | What Makes Knowledge Scientific?

Expressions such as “scientifically proven,” “data-driven,” and “statistically significant” can make a claim sound convincing. Yet using scientific methods does not necessarily make a conclusion correct. Likewise, a claim may employ technical terminology and mathematical formulas without meeting the standards of scientific investigation.

The philosophy of science examines fundamental questions about science: What makes knowledge scientific? How are scientific theories developed and evaluated? How can science be distinguished from pseudoscience? And to what extent do scientific theories describe reality?

Studying philosophy of science will not directly solve most business problems. Nevertheless, a basic understanding of its central ideas can help us avoid naive scientism—the assumption that something must be true simply because it appears scientific—and evaluate theories, evidence, and analytical results more critically.

Karl Popper and Falsifiability

The philosopher Karl Popper (1902–1994) proposed falsifiability as a criterion for distinguishing science from non-science.

Consider the statement “All ravens are black.” Observing ten thousand black ravens cannot establish that every raven is black. Yet discovering a single white raven would refute this universal statement.

Popper argued that scientific theories should not merely accumulate supporting observations. They should also specify what observations would count against them. This approach is known as falsificationism.

For example, a claim that a particular technique always improves creativity could be tested if creativity and the experimental conditions were clearly defined. But if every unsuccessful result were dismissed by claiming that participants lacked sufficient motivation or that “true creativity cannot be measured,” the claim would become difficult to test scientifically.

Falsifiability, however, does not establish whether a theory is true. Nor does a single unexpected observation necessarily invalidate an entire scientific theory. Experimental results also depend on measurement procedures, background assumptions, and auxiliary hypotheses.

Although Popper’s criterion does not provide a complete account of scientific practice, it highlights the importance of looking for evidence that could challenge a theory rather than only evidence that supports it.

Thomas Kuhn and Scientific Paradigms

Does science progress simply by accumulating new facts?

Thomas Kuhn (1922–1996), a historian and philosopher of science, challenged this view in his influential book The Structure of Scientific Revolutions (1962).

Kuhn argued that scientific communities operate within shared frameworks that shape their questions, theories, methods, and standards of acceptable explanation. He called these frameworks paradigms.

During periods of “normal science,” researchers solve problems within an established paradigm. As unexplained anomalies accumulate, confidence in the existing framework may weaken. Eventually, a scientific revolution may replace it with another paradigm.

The transition from geocentric to heliocentric astronomy and the development from Newtonian mechanics to relativity illustrate major changes in scientific frameworks.

Kuhn’s work also suggests that scientific observations are not entirely independent of existing theories. What researchers consider important, how they interpret evidence, and which questions they investigate are influenced by their conceptual frameworks.

This does not mean that every scientific theory is equally valid or that science is merely subjective. Rather, Kuhn emphasized that scientific revolutions can transform the standards used to evaluate competing explanations.

Instrumentalism — Are Scientific Theories Descriptions of Reality or Useful Tools?

Do scientific theories describe the world as it actually exists, or are they primarily instruments for organizing observations and making predictions?

Instrumentalism emphasizes the latter interpretation.

Scientific realism regards scientific theories as attempts to describe the structure of reality, including entities and processes that cannot be directly observed. Instrumentalism focuses instead on their usefulness for organizing experience and predicting phenomena, without necessarily treating every element of a theory as a literal representation of reality.

This distinction is particularly relevant to modern data science.

For example, a machine learning model may accurately predict customer churn without explaining why customers leave. Variables useful for prediction are not necessarily the causes of the predicted outcome.

Similarly, latent variables and clusters derived from multivariate analysis may provide useful representations of patterns in data. Their usefulness does not establish that they correspond to independently existing entities in the real world.

Instrumentalism encourages us to distinguish between a model that works and a model that explains reality. Both objectives can be valuable, but they are not identical.

Occam’s Razor — The Principle of Simplicity

Occam’s Razor is the principle that unnecessary assumptions should be avoided when competing explanations adequately account for the same observations.

It is named after William of Ockham, a fourteenth-century philosopher and theologian who emphasized parsimony in philosophical and theological reasoning.

The familiar formulation that entities should not be multiplied beyond necessity is not found verbatim in Ockham’s own writings. Moreover, the preference for simpler explanations predates him.

The principle remains relevant to contemporary statistics and machine learning.

Suppose we attempt to fit a mathematical curve to ten observations. A sufficiently complex function may fit every observed point perfectly. However, such a model may also capture random noise, leading to poor predictions on new data. This is known as overfitting.

Statistical model-selection criteria such as AIC (Akaike Information Criterion) and BIC (Bayesian Information Criterion), along with regularization techniques in machine learning, address the balance between goodness of fit and model complexity. Although their mathematical foundations and objectives differ, they share a concern with avoiding unnecessary complexity.

Simplicity, however, is not a guarantee of truth. A model can also be too simple to represent important features of reality. The challenge is to achieve adequate explanatory or predictive performance without introducing unnecessary assumptions or complexity.

Distinguishing Science from Pseudoscience

The question of how to distinguish science from pseudoscience is known as the demarcation problem.

Falsifiability is one influential proposed criterion, but no single universally accepted rule provides a complete solution.

It is also important to distinguish non-science from pseudoscience. Art, ethics, religion, and philosophy may pursue objectives that differ from those of empirical science. Being non-scientific does not automatically make something pseudoscientific.

The problem arises when claims that lack adequate scientific support are presented as established scientific knowledge.

When evaluating a claim, several questions are worth considering:

  1. Testability: What observations or measurements could support or contradict the claim?
  2. Quality of Evidence: Is the evidence anecdotal, observational, or based on controlled experiments?
  3. Reproducibility: Can independent researchers obtain comparable results?
  4. Alternative Explanations: Have chance, confounding variables, measurement errors, and selection biases been considered?
  5. Response to Criticism: Is the claim revised when credible contradictory evidence emerges?

These are not infallible rules for classifying science and pseudoscience. They are practical considerations for evaluating the reliability of scientific claims.

Why Philosophy of Science Matters in the AI Era

Generative AI can produce fluent explanations of scientific theories, complex phenomena, and analytical results. However, a plausible explanation is not necessarily a correct explanation.

Asking AI why a particular conclusion was reached may produce a convincing response, but that response is not necessarily a faithful account of the evidence or reasoning underlying the result.

The philosophy of science provides useful questions to ask when working with AI.

What evidence could falsify this hypothesis? Are we interpreting the information through an established paradigm that excludes alternative perspectives? Are we confusing predictive performance with causal explanation? Are we introducing unnecessary assumptions or complexity?

Philosophy of science does not provide definitive answers to every such question. Instead, it offers a foundation for examining the assumptions, evidence, and limitations of scientific claims.

Its central value is not to tell us what to believe, but to help us distinguish what is known, what is assumed, and what remains uncertain.

References and Further Reading