A Nature article presents a language-guided inverse-design framework that uses generative artificial intelligence to design multiclass metamaterials. Metamaterials are architected materials whose effective properties and functions arise primarily from the geometry, connectivity, and spatial organization of their constituent elements, enabling unusual electromagnetic, acoustic, thermal, photonic, and mechanical behavior for applications in robotics, cloaking and wave manipulation, thermal management, optical devices, and catalysis. The strong and often nonlinear coupling between microstructures and effective response makes discovering metamaterials with prescribed properties a long-standing inverse-design challenge. Conventional development follows a forward workflow in which a researcher proposes a microstructure based on physical intuition and prior experience, evaluates its properties through analytical models or numerical simulations, then fabricates and experimentally validates promising candidates, commonly by additive manufacturing; each modification requires another cycle of modeling, simulation, and testing, making the workflow slow and strongly dependent on expert intuition and experience. Recent artificial intelligence and deep learning advances have introduced data-driven alternatives, including multilayer perceptrons and convolutional neural networks as surrogate models that predict effective properties substantially faster than repeated finite-element simulations, and deep generative models such as generative adversarial networks, variational autoencoders, and diffusion models that generate candidate architectures conditioned on target properties. Most existing inverse-design frameworks accept only predefined numerical conditions of materials properties, such as relative density, elastic moduli, Poisson’s ratio, stress–strain responses, thermal conductivity, and wave frequency, which provides precise property control but requires users to understand the model’s