Terahertz (THz) communication is an up-and-coming candidate technology for 6G, aiming to meet the stringent performance requirements of next-generation communications, such as hyper rates on the order of terabits-per-second. However, a notable challenge in THz communications is the increased path losses. Fortunately, extremely large-scale multiple-input multiple-output (XL-MIMO) technology can effectively compensate for this issue. The combined implementation of THz and XL-MIMO, offering extensive spatial degrees of freedom and ultra-wideband capabilities, has emerged as a crucial technology for high-precision target sensing and high-throughput communications, effectively meeting the needs of integrated sensing and communication (ISAC). However, the large antenna aperture of XL-MIMO will make near-field communications dominant. On the other hand, the ultra-wideband nature of THz signals will introduce the beam-split effect. There is an urgent need to consider these unique characteristics to synergize artificial intelligence (AI) with XL-MIMO for developing THz communication and sensing. This proposal aims to design AI-native XL-MIMO THz communication by developing novel techniques: 1) Intelligent THz XL-MIMO Communication Design. Developing low-complexity, learning-aided channel estimation methods for XL-MIMO communication systems while studying intelligent precoding to manage hybrid field interference for mixed near-field and far-field communication users. 2) Hybrid-Field THz XL-MIMO ISAC Design. Developing ISAC precoder for target sensing and electromagnetic property sensing, while developing a network-level ISAC performance analysis framework. 3) Practical implementation of Intelligent THz XL-MIMO ISAC. Developing resilient AI-native XL-MIMO ISAC using generative diffusion model and intelligent XL-MIMO ISAC testbed. The main research approaches include electromagnetic wave theory, nonlinear optimization, and reinforcement learning, etc.