Abstract
<jats:p>Magnetic-material design is a stringent test for artificial intelligence (AI) because the measured response is rarely a property of composition alone. Molecular magnets, spin-crossover compounds, low-dimensional magnets, magnetocaloric phases and permanent magnets connect local electronic structure and spin–orbit coupling to vibrations, crystal packing, defects, morphology, processing and measurement history. This review critically assesses how digital technologies are beginning to navigate that multiscale chain. We distinguish comparatively intrinsic or local targets - spin state, exchange and magnetic anisotropy - from emergent sample-level observables such as blocking temperature, hysteresis, coercivity and magnetocaloric response, and grade the evidence from retrospective prediction to prospective experimental and closed-loop validation. Curated molecular-nanomagnet datasets, high-throughput quantum chemistry, machine learning of spin–phonon dynamics, active learning and human–AI co-expert workflows have already expanded tractable search spaces. However, most studies remain limited by sparse, positive-biased data, inconsistent labels, weak out-of-distribution tests and incomplete capture of raw measurements, failed syntheses, polymorphs and processing history. We therefore argue that the fundamental data object for magnetic AI should be a sample-context record linking chemical identity, structure, preparation, physical state, measurement protocol, raw signal, fitting model and uncertainty. Near-term progress is most likely to come from physics-aware models, leakage-resistant benchmarks, calibrated abstention and modular human-supervised feedback loops. The resulting agenda is not a push-button magnet-design machine, but an auditable infrastructure in which AI helps formulate, rank and test chemically meaningful hypotheses. This review is intentionally focused and selective: representative, field-shaping studies are used to define the current frontier, identify gaps, and propose priorities for trustworthy AI-enabled magnetochemistry.</jats:p>