A machine predicts tomorrow's tide to the centimeter.
Another machine predicts that a molecule will bind to a protein.
A third predicts the trajectory of a physical system it has never seen before.
All three predictions are correct.
Have the machines discovered anything?
The answer is not obvious.
Prediction is one of the great achievements of science. But science has usually wanted something more.
It wants to know why.
Prefer to watch? Continue with the video below, or keep reading.
A Perfect Oracle
Imagine an oracle that can answer any experimental question.
Ask:
What will happen if I double the pressure?
It answers correctly.
Ask:
Which of these molecules will react fastest?
Correct again.
Ask:
What will the system look like in ten seconds?
Perfect.
This oracle would be enormously useful.
Engineering could advance quickly.
Experiments could be planned efficiently.
But if the oracle never provides a mechanism, a principle, or an explanation, would we call it a scientist?
This thought experiment separates two goals that are often mixed together:
\[ \text{prediction} \neq \text{explanation}. \]
Correlation Can Be Enough for Prediction
Machine learning is exceptionally good at discovering statistical regularities.
If pattern $A$ tends to appear before outcome $B$, a model may learn to predict $B$.
For many applications, this is enough.
If a weather model forecasts a hurricane accurately, people care deeply about the forecast even if the internal representation is not a human-readable theory of atmospheric dynamics.
If a medical model detects a dangerous condition early, predictive value matters.
But scientific understanding often asks a different question:
If we intervene on $A$, will \[ \text{umbrella use} \longleftrightarrow \text{wet streets}. \]0 change?
That is a causal question.
Correlation alone cannot always answer it.
The Umbrella Problem
Suppose an AI studies historical data and learns:
\[ \text{umbrella use} \longleftrightarrow \text{wet streets}. \]
The correlation is excellent.
But opening umbrellas does not make streets wet.
Rain causes both.
A predictor can exploit the relationship without understanding the cause.
Science often needs the hidden structure:
\[ \text{rain} \rightarrow \begin{cases} \text{umbrellas}\\ \text{wet streets} \end{cases} \]
This matters because explanation supports intervention.
If we want to change the outcome, we need to know where to act.
Causality Enters the Laboratory
Experiments are powerful partly because they help distinguish correlation from causation.
Instead of merely observing that two quantities vary together, we deliberately change one and watch what happens.
That is an intervention.
This is the middle rung of computer scientist Judea Pearl's Ladder of Causation: moving from association—observing what happens—to intervention—actively doing something and watching the consequences. Pearl's \[ \text{umbrella use} \longleftrightarrow \text{wet streets}. \]1 notation makes the distinction explicit.
In an idealized form:
\[ \text{observe }P(Y\mid X) \]
is not the same as asking
\[ P(Y\mid do(X)). \]
The second represents an intervention on \[ \text{umbrella use} \longleftrightarrow \text{wet streets}. \]2.
The distinction is central to causal inference.
For AI4S, the important point is that an agent connected to a laboratory does not have to remain a passive observer.
It can choose interventions.
That creates a route from prediction toward mechanism.
A Model Can Be Right for the Wrong Reason
Suppose an image classifier recognizes diseased tissue with high accuracy.
Later, researchers discover that it was partially relying on a scanner artifact associated with one hospital.
The predictions were good.
The learned explanation was wrong.
Scientific models face the same danger.
A model may fit observed data beautifully while depending on a spurious relationship that fails under new conditions.
This is why generalization matters.
A useful scientific model should survive when we change the setting, intervene on the system, or test a regime not represented in the training data.
What Counts as Discovery?
There is no single definition.
At least four levels are worth distinguishing.
Prediction
The system accurately forecasts an outcome.
Pattern discovery
The system identifies a regularity that was previously unnoticed.
Mechanistic discovery
The system identifies a causal or structural explanation for the pattern.
Conceptual discovery
The system introduces a new representation, variable, or principle that changes how the problem is understood.
AI has already shown strong capabilities in the first two categories.
The last two are harder.
And they are closer to what people often mean when they speak of major scientific discovery.
AlphaFold Is an Interesting Example
Protein structure prediction is scientifically transformative.
Given sequence information, AlphaFold can predict three-dimensional protein structures with remarkable accuracy across a vast range of proteins.
That capability has changed how researchers work.
But predicting a structure is not identical to discovering every physical mechanism that governs protein folding.
The distinction does not reduce the achievement.
It clarifies it.
A tool can be revolutionary because prediction itself unlocks new science.
Yet explanation remains a separate scientific goal.
This is a useful warning for AI4S generally: an extraordinarily accurate predictor can transform a field without, by itself, supplying the complete causal or mechanistic theory behind its predictions.
Sometimes Prediction Comes Before Theory
History also warns us not to demand too much too soon.
Scientists have often discovered reliable empirical laws before they understood the underlying mechanism.
Astronomical regularities preceded Newtonian mechanics.
Thermodynamics developed powerful principles before a full microscopic statistical picture was established.
A predictive model can therefore be a stage on the path to explanation.
AI may discover patterns first.
Humans—or later AI systems—may then build theories that explain them.
The important question is whether the predictive system helps reveal deeper structure or simply remains a black box.
Abduction: Reasoning Backward from the Observation
There is a form of reasoning especially important here: abduction.
Deduction asks:
Given these laws and conditions, what follows?
Induction asks:
Given many examples, what pattern is likely?
Abduction asks:
Given this surprising observation, what explanation could account for it?
A scientist sees an anomaly and says:
Something must be missing from our model.
That move is central to discovery.
It requires imagination as well as logic.
The explanation is not directly contained in the observation.
It has to be proposed.
Then it must generate new predictions that can be tested.
A powerful scientific AI therefore needs more than pattern recognition.
It needs a way to invent candidate explanations and expose them to evidence.
The Best Experiment May Be the One That Breaks the Model
If our only goal is prediction accuracy, we may keep collecting familiar examples.
If our goal is discovery, we often want experiments where competing explanations disagree.
Suppose Model A and Model B make identical predictions under normal conditions.
But under extreme pressure:
\[ A \rightarrow 5.2, \qquad B \rightarrow 9.7. \]
That is the experiment we should run.
A scientific agent should actively seek such disagreements.
It should look for places where the current model is vulnerable.
That is how prediction becomes a tool for explanation.
A New Relationship Between AI and Theory
One possible future is that AI changes how scientific theories are built.
Instead of beginning with a compact human theory and deriving predictions, researchers may begin with a powerful predictive model.
Then they ask:
- What variables does the model rely on?
- Under what interventions does it fail?
- Can its behavior be compressed into a simpler rule?
- Can symbolic relationships be extracted?
- Which latent features correspond to physical quantities?
- What experiments separate competing interpretations?
In this picture, AI is not the final theory.
It is a machine for generating clues from which theory can emerge.
Discovery Requires More Than Being Right
A prediction can be correct without being explanatory.
An explanation can be plausible without being causal.
A causal model can be useful without being conceptually deep.
Scientific discovery often moves through all of these levels.
So when we hear that AI has “discovered” something, it is worth asking:
What exactly was discovered?
A correlation?
A useful prediction?
A new candidate?
A mechanism?
A new mathematical structure?
A new concept?
These are not the same achievement.
The distinction will become even more important as AI produces scientific results at increasing scale.
And it leads naturally to an even harder question:
If discovery sometimes requires a new way of seeing the problem, can AI invent new scientific concepts?
Next in the series: Can AI Invent New Scientific Concepts?
0 comments