Underdetermination: When Evidence Is Not Enough

Suppose two theories explain every observation we currently possess.

Both fit the data.

Both predict the same next experiment.

But they describe reality differently.

Which one should we believe?

This is the problem of underdetermination.

Evidence can constrain theories without always selecting one uniquely.

The problem matters because science aims not only to fit data but often to tell us what the world is like.

If multiple incompatible pictures fit equally well, evidence may leave ontology unsettled.

What Underdetermination Means

A theory is underdetermined by evidence when more than one theoretical account is compatible with the available observations.

In simplified form:

Evidence E fits Theory A.

Evidence E also fits Theory B.

Yet A and B differ significantly.

Therefore E alone does not determine which theory is correct.

This sounds abstract.

It occurs throughout science.

Curve Fitting

Imagine five data points.

Many mathematical curves can pass through all five.

A straight line may fit approximately.

A polynomial can fit exactly.

A more complicated function can fit exactly too.

Observed points alone do not uniquely specify the underlying rule.

We choose among models using additional criteria.

This is a simple form of underdetermination.

Finite Data, Infinite Possibilities

Any finite dataset can usually be generated by many possible rules.

Suppose a sequence is:

2, 4, 6, 8.

What comes next?

10 seems natural.

But a contrived rule could produce 1,000 next while matching every previous term.

Evidence constrains possible continuations.

It rarely determines one continuation by logic alone.

Goodman Reappears

This resembles Goodman’s “grue” problem.

Past evidence can support multiple generalizations depending on which predicates we use.

The problem is not just lack of data.

It is that generalization depends on representational choices.

Theory selection therefore requires more than raw observation.

Empirically Equivalent Theories

Two theories are empirically equivalent if they produce the same observable predictions over some domain.

If equivalence were complete and permanent, no experiment within that domain could distinguish them.

Theories might still differ in ontology.

This creates a direct challenge to scientific realism.

Quantum Interpretations

Quantum mechanics provides a familiar example.

Different interpretations can agree on ordinary experimental predictions while disagreeing radically about what exists.

Does the wavefunction collapse?

Do all outcomes occur in branching worlds?

Are there hidden particle trajectories?

If predictions are identical, experimental evidence alone cannot yet choose.

That is underdetermination.

Gauge Freedom Is Not Always Underdetermination

Care is needed.

Sometimes two mathematical descriptions are merely different representations of the same physical state.

Different gauge choices can describe equivalent physics.

Different coordinates can describe the same spacetime.

That is not necessarily competing ontology.

True underdetermination requires theories that differ physically, not merely notationally.

Geocentric and Heliocentric Descriptions

Historical astronomy is often simplified as a battle between two coordinate systems.

But the scientific transition involved more than choosing where to place the origin.

Dynamics, simplicity, telescopic evidence, and Newtonian mechanics changed the explanatory framework.

Equivalent coordinate descriptions should not be confused with equivalent physical theories.

Representation and ontology must be separated.

Duhem’s Problem

Pierre Duhem emphasized that experiments test networks of assumptions.

A prediction depends on:

• central theory,
• auxiliary hypotheses,
• instrument models,
• initial conditions.

When evidence conflicts with prediction, many revisions are possible.

Evidence may therefore underdetermine which part of the theoretical network should change.

Quine’s Extension

W. V. O. Quine generalized this idea.

Our beliefs confront experience as an interconnected web.

In principle, many parts of the web can be adjusted to accommodate recalcitrant evidence.

This is sometimes called confirmation holism.

Scientific testing is therefore less atomistic than simple textbook examples suggest.

Strong vs Weak Underdetermination

A useful distinction is between temporary and permanent underdetermination.

Temporary

Current evidence cannot distinguish theories, but future experiments may.

Permanent

The theories are constructed so that no possible observation can distinguish them.

Temporary underdetermination is routine.

Permanent underdetermination is philosophically more troubling.

Future Evidence

History shows that apparently equivalent theories can later diverge.

Better instruments reveal differences.

New domains become accessible.

A theory makes an unexpected prediction.

So present underdetermination does not imply permanent equivalence.

Science often resolves ambiguity by expanding observation.

Theory Virtues

When evidence alone does not decide, scientists use additional criteria.

These may include:

• simplicity,
• explanatory scope,
• coherence,
• unification,
• fertility,
• compatibility with established knowledge,
• computational tractability.

These are sometimes called theoretical virtues.

But are they evidence for truth or merely pragmatic preferences?

That question is controversial.

Simplicity Again

Ockham’s razor is one response.

If two theories fit equally well, prefer the one with fewer unsupported assumptions.

But simplicity may depend on representation.

A theory simple in one language can look complicated in another.

Simplicity helps.

It does not mechanically solve underdetermination.

Explanatory Power

Theory A may merely fit observations.

Theory B may explain why the pattern occurs through a mechanism.

Scientists often prefer B.

But if both predict identically, is explanatory satisfaction evidence of truth?

Realists often say yes, at least indirectly.

Instrumentalists are more cautious.

Unification

A theory that explains several previously disconnected domains with one framework gains credibility.

Maxwell’s theory unified electricity, magnetism, and light.

General relativity unified gravitational effects through spacetime geometry.

Unification reduces independent assumptions.

It can break practical ties among empirically adequate models.

Novel Predictions

The strongest way to resolve underdetermination is often to find a new test.

Two theories agree on old data.

Search for a case where they diverge.

Then perform the experiment.

Scientific creativity often consists of finding the discriminating observation.

Auxiliary Hypotheses

Suppose Theory A fails a prediction.

Its supporters modify an auxiliary assumption.

Now it fits again.

Theory B does the same with a different adjustment.

At what point does one research programme become too ad hoc?

There is no algorithmic answer.

Scientists judge whether modifications produce independent explanatory gains.

Lakatos and Progressive Programmes

Imre Lakatos suggested comparing research programmes over time.

A progressive programme predicts novel facts successfully.

A degenerating programme mainly repairs itself after failures.

This turns theory choice into a historical pattern rather than a single experiment.

Underdetermination may be reduced by looking at long-term scientific productivity.

Bayesian Model Comparison

Bayesian methods provide a formal framework.

Each theory receives a prior probability.

Evidence updates the probabilities according to likelihood.

If two theories predict the evidence equally well, their posterior ratio depends on priors.

New discriminating evidence can shift the balance.

Bayesianism does not eliminate underdetermination.

It makes assumptions explicit.

Priors Create Their Own Problem

Where should prior probabilities come from?

Simplicity?

Symmetry?

Past success?

Subjective judgment?

Different reasonable priors can produce different conclusions when evidence is weak.

As evidence accumulates, priors often matter less.

At the frontier, they can matter greatly.

Machine Learning Analogy

Machine learning faces underdetermination constantly.

Many models can fit training data.

The challenge is generalization.

Researchers use:

• regularization,
• validation sets,
• inductive bias,
• architecture choices.

These are computational versions of theory-choice principles.

The problem is not unique to philosophy.

It is built into learning from finite data.

Inductive Bias

A learner cannot generalize without assumptions.

Prefer smooth functions.

Prefer simpler models.

Prefer local continuity.

Prefer certain symmetries.

These assumptions are inductive biases.

Science also has inductive biases.

We prefer theories with regularity, coherence, and simplicity because without such preferences, finite evidence cannot guide us efficiently.

Does This Mean Science Is Arbitrary?

No.

Underdetermination does not imply that any theory is as good as any other.

Evidence can rule out enormous regions of possibility.

Some theories fit poorly.

Some require ad hoc rescue.

Some conflict with independent knowledge.

Some fail novel predictions.

The problem is that surviving evidence may leave more than one serious option.

Constraint is not uniqueness.

Does Consensus Solve It?

Scientific communities often converge despite underdetermination.

Why?

Because theories differ in many dimensions beyond raw fit.

Researchers evaluate:

• elegance,
• mechanism,
• tractability,
• compatibility,
• future research potential.

Consensus can therefore emerge before decisive proof.

This is rational, but still fallible.

Social Factors

Theory choice also occurs inside communities.

Education.

Prestige.

Funding.

Research traditions.

Available instruments.

These influence which theories receive attention.

Recognizing social factors does not imply truth is socially constructed in an unrestricted sense.

Reality still constrains outcomes.

But scientific pathways are historically situated.

Realism Under Pressure

Underdetermination challenges realism because evidence may support several incompatible ontologies.

A realist must explain why one deserves belief.

Possible responses include:

• future evidence will decide,
• theoretical virtues track truth,
• only shared structure deserves realist commitment.

This is one reason structural realism is attractive.

Instrumentalist Response

An instrumentalist can say:

If two theories make exactly the same predictions, there is no scientific need to decide which hidden ontology is “really there.”

Use whichever model is more convenient.

This dissolves some of the problem.

But many scientists still want explanation, not only prediction.

The philosophical tension remains.

Underdetermination in Cosmology

Cosmology is especially vulnerable.

We observe one universe.

We cannot rerun its initial conditions.

Several early-universe models may produce similar present signatures.

This is why claims about:

• pre-Big-Bang phases,
• inflationary mechanisms,
• multiverses

require caution.

The farther the inference from observable evidence, the more room alternative models may have.

Underdetermination in Consciousness

Later, the problem will return.

The same behavioral and neural evidence may fit different theories of consciousness.

Functionalism.

Global workspace.

Integrated information.

Illusionism.

Panpsychism.

Some differences may be empirically testable.

Others may remain partly philosophical.

Underdetermination is not confined to physics.

The Scientific Strategy

When evidence is not enough, good science does not surrender.

It asks:

What new observation would distinguish the theories?

Can assumptions be reduced?

Can one theory explain more?

Can the models be connected to independent evidence?

Can hidden differences become measurable?

Underdetermination becomes a research programme.

The Next Problem

Science does not only compare static theories.

It changes historically.

Entire frameworks rise, dominate, face anomalies, and sometimes collapse.

Thomas Kuhn argued that such changes are not simply smooth accumulation.

They can be scientific revolutions.

The next question is:

How do paradigms shape science, and what happens when one paradigm replaces another?