AI Can Now Turn a Chart Screenshot Into Working Interactive Code

Charts on websites aren't just pictures. They're small programs. Hover over a bar and a number pops up; click a dropdown and the chart redraws itself. A team of researchers has released the first serious test of whether AI can do that whole job: look at a screenshot of a chart and write the code that rebuilds it — data, colors, and interactions included. The test, called IChart2Code, contains 377 chart tasks across 20 chart styles and 13 kinds of data, with 1,209 specific interactive requirements built in.

The results are a reality check. Today's best AI models — the kind that can already "see" images and write code — handle the easy part reasonably well: matching how a static chart looks. Then they fall down. They get the underlying numbers wrong, misplace labels, and often can't make the interactivity work once the code actually runs in a web browser. The researchers judged every attempt inside a real browser, so a chart that looks right but breaks when you click it earns no credit.

To help, the team built a tool called TRAIL. Think of it as giving the AI a proofreader: a second AI "Inspector" writes a checklist, runs the code in a browser, watches what happens, and hands back specific repair notes — "the tooltip shows up in the wrong corner," that kind of thing. Averaged across four different AI models, TRAIL lifted scores by roughly 3 to 8 percentage points depending on the category. That's real progress, but not a leap.

Why should you care? Building an interactive chart is fiddly, tedious work — the sort of thing an analyst, marketer, or data journalist might lose an afternoon to. If AI can hand back a rough first draft in seconds, that's genuine time returned to you. But because the models still get numbers and interactions wrong, a person has to double-check the output. And a chart with a wrong number is far worse than a chart that's merely ugly, because it looks so authoritative.