Optical neural networks, which use photons as information carriers and are used to construct optical neurons and optical synaptic connection models, are key components for achieving high-performance computing tasks in optical artificial intelligence. Metasurfaces, composed of subwavelength artificial nanostructures capable of multidimensional light field manipulation, have become a new platform for optical neural networks, offering new possibilities for parallel processing of large-scale intensive computational tasks. This article presents a review of the development and challenges of metasurface-based optical diffraction neural networks. First, it introduces the realization of optical field multitask parallel-processing approaches using metasurface optical diffraction neural networks, the preparation of chip integration, and their applications in image recognition and classification, optical information processing, and complex visual tasks. Finally, it concludes with an outlook on the challenges and future development of metasurface optical diffraction neural networks.