How to Test a Brain Explanation That Sounds Like It Explains Everything

Grand theories of the brain are hard to resist.

There is something satisfying about finding one idea that seems to connect perception, learning, action, hallucinations, emotion, and maybe even consciousness.

Predictive processing can feel like that.

Once you learn the vocabulary—priors, prediction errors, precision—many very different phenomena suddenly seem to fit into the same picture.

That is useful.

It is also exactly when a second question becomes important:

If a theory can explain almost anything, how do we know when it has actually explained something?

A Prediction That Cannot Lose

Imagine a weather forecaster saying:

“Tomorrow it might be sunny, rainy, windy, or snowy.”

Whatever happens, the forecast survives.

But surviving is not the same as being informative.

A useful prediction rules things out. Saying there will be heavy rain tomorrow afternoon is much riskier because the world can prove you wrong.

Scientific explanations work in much the same way.

The fact that a theory can accommodate a result after it happens is useful, but weak. Stronger evidence appears when the theory commits to something in advance—especially when another plausible explanation predicts something different.

“Consistent With” Is Only the Beginning

Consider Mismatch Negativity, or MMN^[1].

Play the same tone repeatedly—pip, pip, pip, pip—and then suddenly replace one with a different sound.

The deviant tone produces a characteristic neural response.

Predictive coding offers an elegant interpretation: the repeated tones establish an expectation, the deviant violates it, and the resulting mismatch contributes to updating the model^[1].

That explanation fits well.

But it is not the only one.

Neurons responding to the repeated sound may have adapted, leaving neurons tuned to the rare sound relatively fresh. The unusual tone may also capture attention simply because it is rare.

Those mechanisms can also produce a stronger response to the deviant.

So MMN is certainly compatible with predictive coding. What it does not do, by itself, is uniquely identify predictive coding as the mechanism.

This distinction matters far beyond MMN.

The useful question is not merely:

Can my theory explain this result?

It is:

What did my theory predict that the alternatives did not?

Getting Closer to the Mechanism

Some experiments make that question easier to answer.

In a well-known study, mice moved through a virtual environment while researchers recorded activity in primary visual cortex^[2].

Normally, running produced matching visual flow. Then, on some trials, the visual flow was suddenly disrupted while the mouse continued moving.

Certain neurons in superficial layers of visual cortex responded strongly to that mismatch^[2].

The interesting part is that the response was tied to the relationship between movement and visual feedback, not simply to visual motion itself.

That makes the finding more mechanistically informative than a generic “unexpected things produce more brain activity” result.

Still, it is worth keeping the conclusion narrow.

The experiment provides strong evidence that mouse visual cortex contains neurons sensitive to sensorimotor mismatch. It does not, on its own, prove that every cortical area implements the same predictive-coding architecture—or that the same mechanism explains high-level human thought.

A Framework Is Not the Same as a Mechanism

This is where broad frameworks and specific models need to be separated^[3].

A framework gives researchers a way to organize a problem.

Predictive processing, for example, suggests thinking about perception in terms of expectations, incoming evidence, uncertainty, and mismatch.

That can generate useful questions even before we know the exact biological implementation.

A process model goes further.

It might propose that particular cortical layers carry different signals, with some pathways emphasizing predictions and others mismatch information^[4].

Now the theory has made itself more vulnerable.

Researchers can ask whether those layers actually behave as predicted, whether disrupting one pathway changes the expected signal, and whether a competing architecture explains the data better.

That vulnerability is a strength.

A model becomes more scientifically useful as it becomes clearer what would make it fail.

The Precision Problem

Predictive explanations often rely on another useful concept: precision.

Roughly speaking, precision describes how much confidence the system places in different sources of information.

A sensory signal judged highly reliable should influence perception more strongly than one judged noisy or uncertain.

The idea is powerful.

It can also become dangerously convenient.

Suppose someone ignores an obvious signal. We could say that sensory precision was low.

If another person responds strongly to the same signal, we could say their sensory precision was high.

Both explanations may be reasonable.

But if the behavior itself is the only reason we concluded that precision was high or low, and we then use that inferred precision to explain the same behavior, we have gone in a circle.

Latent variables are not a problem by themselves. Science uses unobserved variables all the time.

The important question is whether they are constrained by something beyond the result they are supposed to explain—another measurement, an experimental manipulation, a preregistered model, or successful prediction of new data.

Four Questions Worth Asking

You do not need to be a neuroscientist to spot many of these problems.

When you encounter an explanation of the brain, four questions go a long way.

1. What did the model predict before the result was known?
Was the outcome genuinely anticipated, or was the story assembled afterwards?

2. Could another plausible explanation produce the same result?
A finding that fits several models does not strongly distinguish between them.

3. Are the hidden parameters constrained independently?
Are things like precision, priors, or uncertainty estimated from additional evidence, or introduced only because they make the result fit?

4. What result would count against this specific model?
If the answer is “nothing,” the claim is probably too vague to test properly.

Why This Does Not Kill Predictive Processing

None of this means predictive processing should be discarded.

Quite the opposite.

Its value is not that it gives us a phrase that can be attached to every phenomenon. Its value is that it can generate increasingly specific models about what the nervous system should do.

Some of those models will work.

Some will fail.

Some findings will turn out to have simpler explanations.

That is not a weakness of science. That is the process that turns an attractive framework into a better theory of the brain.

After seven articles asking what prediction might explain, the final question is therefore a different one:

What would have to happen for this explanation to be wrong?

If we can answer that clearly, we have moved beyond a good story.

We have something we can actually test.