Every generation invents more powerful tools. Every generation also risks mistaking more powerful tools for deeper understanding.
Twenty-three years ago I wrote a Nature essay asking whether increasingly realistic computer graphics could become misleading. My concern was not that the images were inaccurate. It was that realism could be mistaken for explanation. A compelling visualization might persuade us that we understood a phenomenon when, in reality, we had simply produced a more convincing picture.
Today I find myself making the same argument—but this time about computation and AI.
Our recent review in Applied Mechanics Reviews examines granular segregation, a phenomenon that can now be simulated with extraordinary fidelity. Modern computational methods reproduce the behavior of millions of particles. The predictions are remarkable.
But reproducing a phenomenon is not the same as understanding it.
A simulation tells us what happens. A conceptual model tells us why. Throughout science, conceptual models—from Bohr’s atom to Thomas Schelling’s model of segregation in cities—have revealed essential mechanisms by stripping away unnecessary detail. Their power lies not in realism but in explanation.
Artificial intelligence raises the stakes. AI is becoming extraordinarily good at prediction—forecasting, classifying, optimizing, and recommending—often without providing mechanistic explanations. That is an extraordinary achievement, but it also risks widening the gap between prediction and understanding.
The history of science repeatedly reminds us that advances in representation often outpace advances in understanding. Better tools amplify what we can compute and visualize. They do not, by themselves, deepen understanding.
As our tools become more powerful, conceptual models become not less important, but more important. They remain the bridge between computation and comprehension.
