Researchers at MIT’s Laboratory for Information and Decision Systems, alongside colleagues at Harvard, are asking a simple yet sharp question: when AI models make accurate predictions, do they truly understand the deeper world behind them? What they want to know is whether these predictive systems have built models of the world that let them generalize beyond their training data, reports MIT News.
The team drew a clever parallel to Kepler and Newton. Kepler nailed planetary motion predictions, but only Newton understood why. The researchers are testing AI to see if it operates at the Kepler level, good at specific tasks, or at the Newton level with underlying comprehension.
To measure that, they developed a new metric called inductive bias. It quantifies whether a model’s predictions align with an actual simulation of a system. At first, with simple one-dimensional lattices, think of a frog hopping among lily pads, the predictive systems reconstruct the system quite well. But ramp up complexity, i.e., more dimensions, and the models fail to capture the true structure.
In another test involving the board game Othello, models could spot legal moves, but badly misjudged the true arrangement of pieces on the board—even those temporarily blocked from play. That shows the models grasp immediate predictions but not the bigger setup.
They also evaluated five real-world predictive model classes. Across the board, the more complex the scenario, the more models drifted from real-world logic.
The researchers presented these results at the International Conference on Machine Learning. They hope to offer a test bed where we can evaluate whether AI models truly understand systems or merely mirror patterns. That’s critical in fields such as drug discovery or protein prediction, where false confidence could cost time, money, or worse.