Multi-task learning in AI involves building a general model that can be applied to multiple tasks and challenges, resulting in better performance and representations for each individual task. This approach involves a core model trained on specific tasks, as well as several heads trained on different tasks simultaneously for improved embeddings and representations.