Foundation interatomic potentials learn the practical chemistry they were never taught

Chemists understand molecular reactivity through simplified concepts such as substituent effects, acidity, and aromaticity, which are compact descriptors of the potential energy surface and of the electronic structure that shapes it. Universal machine-learning interatomic potentials (MLIPs) are transforming applied computational chemistry by learning the same surface from atomic structures, energies, and forces, making atomistic simulations of complex systems practical. Like chemists, MLIPs build compact internal representations of that surface, yet these are usually treated as a means to faster, larger-scale molecular dynamics simulations. Here we show that they also contain directly recoverable chemical information: simple models applied to the invariant per-atom embeddings of pretrained universal MLIPs recover Hammett substituent constants, Brønsted and Lewis acidity, and ring aromaticity with Spearman ρ = 0.8–0.99. The same representations predict the yield of a four-component palladium-catalyzed cross-coupling at ρ = 0.96, matching models built from curated descriptors, while pinpointing the aryl halide as the reaction’s key substrate and ranking isoxazole additives by catalyst-poisoning strength. Universal MLIPs are therefore not only fast engines for energies and forces but also chemically meaningful representations of molecular reactivity, opening a route to extracting chemical insights directly from the learned potentials themselves.