Diffusion models in drug discovery are increasingly used to generate new molecular structures. A study by Andrea Mastropietro and Jürgen Bajorath now shows that, in linker design, these AI models primarily rely on geometric patterns rather than systematically capturing chemical relationships.