The quality of AI-generated website design is rapidly improving, making it increasingly difficult to distinguish from human-created work. This advancement is driven by the AI industry's focus on enhancing creative output through refined evaluation metrics and tools. A key challenge in this evolution is teaching AI a sense of "taste," moving beyond mere correctness. Startups and major AI labs are now working to encode subjective aesthetic preferences into models.Developing AI taste involves gathering extensive data and, crucially, learning from human experts. Companies are moving beyond general internet scraping to collect domain-specific, expert-level data sets. This includes curated examples of "best-in-class" designs and vast visual libraries to build a richer vocabulary for AI.The process of teaching AI taste relies heavily on human feedback, particularly through reinforcement learning with human feedback (RLHF). This involves panels of experts rating and commenting on AI outputs to guide models toward desirable outcomes. The subjective nature of taste means that human decisions at every stage, from data selection to feedback interpretation, shape the AI's aesthetic.Ultimately, AI taste is a synthesis of numerous human aesthetic choices. This collaborative approach is leading to new job categories for creative professionals who provide expert feedback. Companies are investing in infrastructure and tools to facilitate this annotation and evaluation process for AI models.