New AI Fix Stops Research Tools From Making Up Chart Numbers

More and more people use AI assistants to research a topic and hand back a report with charts in it — bar graphs, trend lines, comparisons. The problem is that these tools often decide what the chart should look like before they know what data actually exists. So an AI might plan a chart comparing five countries over ten years, then only find data for three countries over four years. Instead of fixing the chart, it sometimes just fills in the gaps with numbers that look real but aren't backed by anything.

A team of researchers from China and the Netherlands propose a fix called Frame-Evidence Co-Adaptation, or FECA. The idea is simple to describe: let the chart and the evidence talk to each other. The AI sketches a possible chart, goes looking for data, and then decides — based on what it actually found — whether to keep that chart, change it, or throw it out and try a different one. It's the difference between a journalist who writes the headline first and one who reports first.

They tested it on 100 real-world research topics. Charts came out with much more accurate numbers, and the overall reports were just as good and just as useful. Notably, the method doesn't require a bigger or more expensive AI model — it changes how the AI plans its work.

Why should you care? AI is increasingly writing the summaries, dashboards, and slide decks people make decisions with. A chart that looks authoritative but rests on invented numbers is worse than no chart at all — it can send a business plan, a school project, or a news story in the wrong direction. This research is an early step toward AI that admits what it doesn't know instead of quietly making it up. It's a preprint, so it hasn't yet been reviewed by other scientists or built into the tools you use.