Imagine asking an AI to create a chart from your data. Sounds simple enough, right? A new research paper, "VisEval: A Benchmark for Data Visualization in the Era of Large Language Models," reveals that it's trickier than it seems. Large Language Models (LLMs) are now being used for all sorts of tasks, including turning data into visuals. However, researchers found that these LLMs often stumble when asked to make charts. The research introduces VisEval, a new test designed to assess how well LLMs can handle data visualization. This test uses a large collection of realistic queries and databases to push LLMs to their limits. The results? While some LLMs were better than others, all of them struggled with certain chart types, especially those requiring complex layering of data (think stacked bar charts). Surprisingly, giving the LLMs more information didn't always help; sometimes it even made things worse. The research also found that even when LLMs successfully created a chart, they sometimes produced visuals that were hard to read—things like overlapping text or poorly chosen scales. Why is this happening? One reason is that LLMs don't fully understand the data they're working with. They might misinterpret a data column or use the wrong type of chart for the given data. This research is a big step forward in understanding the strengths and weaknesses of LLMs for data visualization. It highlights the need for better tools and techniques to help AI not just create charts, but create meaningful and readable charts. The future of data visualization might involve AI, but we need to teach it to see data more clearly first.