AI models struggle to verify facts in interactive charts.

Computer models that read text and images usually look at still pictures. But real data charts are interactive. Information might hide behind a click, show up across linked displays, or appear only when a user explores. To fix this gap, researchers built a testing ground named InSight. They took statements from real human data stories and tied them to working, interactive websites. In this test, an artificial intelligence agent must explore the web environment to check whether each statement is true, false, or impossible to verify. The benchmark also tracks the exact clicks and moves the models make to see how they search for clues. The researchers tested leading models and found that checking facts across interactive charts is still difficult.

Why it matters

For people who use data visualizations to make decisions, AI needs to handle dynamic, interactive environments to be reliable. This benchmark shows current models aren't yet capable of that.

Questions answered