Table-filling AI learned physics patterns but can't handle exact rules or units.

Most science experiments record measurements in simple tables. New artificial intelligence tools called tabular foundation models learn to predict missing table data, similar to how chatbots predict text. Because these models make predictions using built-in assumptions called priors, researchers tested whether they actually understand physics. They tested four of these models against six baseline systems across datasets created from 316 physical equations, checking how they work on familiar and completely new problems. The models beat the baselines easily, both immediately and after extra tuning. However, the researchers discovered that the models cannot handle real physical units like meters or seconds, and their built-in assumptions cannot handle pure, noiseless rules. They can smoothly fill in gaps between physical data points, but cannot truly work as genuine physics models.

Why it matters

Scientists and engineers who use data tables to model physical systems might rely on these AI models for predictions. However, since the models don't understand units or exact mechanisms, their results could be misleading in applications like engineering or climate science.

Questions answered