Imagine asking an AI to design the perfect chart for your data. Intriguing, right? That’s the fascinating question explored by a recent research paper that dives into how Large Language Models (LLMs), like those powering ChatGPT, make visualization design choices. The researchers developed a clever method called DracoGPT to uncover what design principles LLMs have learned by analyzing how they rank existing chart designs and how they create new ones from scratch. Interestingly, they found that while LLMs are pretty consistent in their design preferences, they often stray from best practices established through human testing, particularly when it comes to tasks involving summaries or aggregates. For tasks like comparing values, LLMs showed more agreement with human preferences, hinting that they’ve picked up existing visualization knowledge prevalent in their training data. This opens exciting avenues for future research: how can we refine LLMs to align better with human perception and create truly insightful visualizations? And, can these models eventually become fast, efficient visualization design tools, replacing more cumbersome methods?