Can Science Justify Itself?

On the surface, the answer seems obvious: science works. Airplanes fly, antibiotics cure infections, and large language models autocomplete our thoughts. But under the hood there is a deep and unresolved problem: the very pattern of reasoning that makes science possible - induction - does not appear to have a non-circular justification.¹

This post sketches the problem, connects it to observation and cognition, and surveys a few prominent responses.

What Is Induction, Exactly?

Induction is the move from observed cases to unobserved ones:

The sun has risen every day in recorded history, so it will rise tomorrow.
Every sampled piece of copper conducts electricity, so all copper conducts.
Repeated trials show a drug helps patients, so it will probably help the next one.

Inductive reasoning underlies almost every empirical inference in science: generalizing from experiments, extrapolating trends, and treating unobserved cases as similar to observed ones.

David Hume gave the classic formulation of the problem in the 18th century.² He argued that such inferences seem to assume a Principle of the Uniformity of Nature (PUN): roughly, that the future will resemble the past or that similar causes will have similar effects. But that principle is neither:

Deductively provable: its denial is not a logical contradiction.
Inductively provable: any attempt to support it by past regularities would already presuppose that such regularities are evidence about the future, which is exactly what is at issue.

So if scientific reasoning is essentially inductive, there appears to be no way to show that it is rational without using the very pattern of reasoning being questioned.

Observation Is Not Neutral Either

Induction is not the only fragile piece. Even “raw data” is not as raw as it looks.

Philosophers of science like Norwood Russell Hanson, Thomas Kuhn, and Paul Feyerabend argued that observation is theory-laden.³ Rough idea:

What scientists see (in the sense of “see that this is a track of an electron”, or “see this as a galaxy rotation curve”) is shaped by their existing concepts, models, and expectations.
There is no completely neutral, theory-free observational language that all parties can use to adjudicate between competing theories.

Contemporary work in cognitive psychology largely supports a moderate version of this: top-down expectations can influence perception and interpretation, especially when sensory input is weak or ambiguous.⁴ This reinforces a second limit:

Science tests hypotheses “against observation,” but observation itself is filtered through fallible, theory-infused human cognition.

So we get a double bind: the inference from data is inductive and apparently unjustified, and the data are not clean, objective givens either.

Hume’s Challenge in a Nutshell

Hume’s argument can be distilled into a few key moves:

There are (roughly) two kinds of arguments:
- Demonstrative (deductive): if the premises are true, the conclusion must be true.
- Probable (inductive): the premises make the conclusion more likely, but do not guarantee it.
The claim “the future will resemble the past” (the Uniformity Principle) is not demonstratively certain; its negation is not a contradiction.
Any attempt to justify inductive inferences using experience (for example, “induction has worked well in the past”) already presupposes that past success is a good guide to future success, which is itself an inductive assumption.

So either:

Induction has no rational justification, or
Its justification is circular or question-begging.

This is the problem of induction: not that induction fails in practice, but that its use appears to outrun what can be justified by strict logic or non-circular argument.

Popper, Falsification, and Why That Isn’t a Free Pass

Karl Popper famously tried to sidestep induction by recasting science as a purely deductive enterprise.⁵ On his view:

Science does not confirm theories; it only refutes them.
We propose bold conjectures and use observation to try to falsify them.
When a theory survives attempted falsification, it is “corroborated” but never justified in the inductive sense.

This picture has a certain elegance: it avoids saying that accumulating confirmations make a theory more probably true, and so appears to dodge Hume’s challenge. But there are well-known problems:

Scientists do treat surviving prediction tests as some evidence for a theory; this looks like inductive support, even if one avoids the word “justification.”
Falsification itself depends on auxiliary assumptions (about instruments, background theories, ceteris paribus clauses); deciding when a theory is really refuted often involves inductive judgment about which assumptions to retain or abandon.⁶

Popper’s legacy is still huge - especially the focus on risky predictions and refutation - but most philosophers think his view does not fully dissolve the problem of induction, it just rephrases it.

Reliabilism and Pragmatism: “Works” as Justification?

More recent approaches attack the problem by changing what “justified” means.

Reliabilism. Reliabilist epistemology says, roughly: a belief is justified if it is produced by a reliable process, i.e., one that tends to generate a high ratio of true to false beliefs.⁷

Frank Ramsey is often cited as an early reliabilist: he suggested that reasonable beliefs are those arising from truth-conducive “mental habits,” evaluated pragmatically by how well they guide successful action.⁸ Alvin Goldman and others later formalized this into process reliabilism, where:

What matters is not an internal proof that a method is trustworthy, but the actual performance of that method in the world.
Inductive practices (statistical inference, controlled experiment, etc.) can be epistemically good simply because, in our world, they reliably track regularities.

This does not give a Hume-style non-circular proof from the armchair; it instead says that objective reliability is enough for justification, even if agents cannot demonstrate it without circularity.

Pragmatic justification. A related move is explicitly pragmatic:

If any method will let us predict, control, and navigate our environment, it looks a lot like induction and modern scientific method.
Alternatives (pure guesswork, arbitrary rules, radical skepticism) perform dramatically worse by our own standards of success.

On this view, even if we cannot prove that inductive rules will continue to work, it is rational - given our aims of survival, prediction, and explanation - to adopt them rather than stand still. The justification is not “these rules mirror the necessary structure of reality,” but “given what we care about, and what we have seen, using them is our best option.”⁹

You can think of this as elevating “works well in practice” from a weak heuristic to a central normative criterion.

Theory-Ladenness and Cognitive Limits

The worries about induction connect naturally with worries about cognitive bias and theory-ladenness of observation.

Cognitive science documents systematic biases: confirmation bias, availability heuristics, motivated reasoning, and more, all of which can distort scientific judgment.
Theory-ladenness suggests that what counts as an “observation” is partly determined by prior theoretical commitments and conceptual schemes.

These issues amplify Hume’s challenge:

Even if induction had a clean logical status, the inputs to inductive inference - our observations and data sets - are filtered through noisy perceptual systems, social structures, and background theories.

Contemporary responses emphasize methodological safeguards - randomization, blinding, preregistration, replication, triangulation across methods - to reduce the impact of individual cognitive limitations. But those safeguards themselves are justified inductively: we use them because they have historically reduced error rates.