Machine-learned interatomic potentials truncate electrostatic interactions at a cutoff radius. Beyond the cutoff, atoms don't see each other. The approximation seems reasonable — long-range interactions are weak, and including them is expensive. The error should be quantitative: slightly wrong energies, slightly wrong forces.
Parker et al. (arXiv:2603.04228) show the error is qualitative. Water at interfaces modeled with short-range ML potentials undergoes a spurious phase transition. The dipole fluctuations grow large enough to look like electrical conductivity. The model predicts metallic water — an entirely different phase of matter that doesn't exist under the simulated conditions.
The mechanism: electrostatic truncation eliminates the long-range correlations that suppress large-amplitude collective dipole fluctuations. In bulk water, these correlations enforce the dielectric screening that keeps water insulating. At interfaces, where the dielectric environment is asymmetric, the missing long-range interactions can't provide the restoring force. The fluctuations grow unchecked. The system crosses a phase boundary that exists only in the model.
The artifact doesn't announce itself. The simulation runs normally. The energy converges. The structure looks reasonable. The phase transition happens smoothly. Nothing in the standard diagnostic toolkit flags the metallization as unphysical. You have to know that water at this interface shouldn't be metallic — which requires external knowledge the model doesn't possess.
The general pattern: a seemingly reasonable approximation (short-range cutoff) doesn't degrade accuracy smoothly. It creates a catastrophic ontological error — the model doesn't just get the number wrong, it gets the phase of matter wrong. The error is not in the last decimal place. It is in the first word.