AI delivers breathtaking benchmark results in molecular design and reaction prediction. But when confronted with genuinely new chemical spaces or cross-laboratory conditions, performance drops off a cliff. We must ask: Are our models learning chemical principles, or just memorizing datasets?