Study: AI's Fancy Graph Explanations Can Actually Make Your Decisions Worse

When AI helps you make decisions — like choosing a medical treatment or reviewing a job candidate — it often explains its reasoning in long blocks of text. But those text explanations can be overwhelming, which is why some researchers think showing AI's logic as a visual graph might be easier to follow. A new study from MIT and Peking University tested that idea with a tool called Graphionale, which turns AI's step-by-step reasoning into a flowchart-style diagram with connected bubbles for conclusions, evidence, and counterpoints.

The researchers ran a large online experiment with 204 people, comparing text explanations versus graph explanations across two types of tasks: verbal reasoning (think reading a contract and spotting problems) and visual reasoning (think analyzing a chart or image). The results were surprising. For verbal tasks, graphs actually helped people calibrate their trust — meaning people trusted the AI more only when the AI was right. But those same people found the graphs more mentally tiring and less satisfying than plain text.

For visual tasks, the opposite happened. Graphs made trust calibration worse — people were more likely to trust a wrong AI — yet they reported feeling more engaged and found the graphs helpful. In both cases, the format people preferred was not the one that gave them the most accurate trust. This "like vs. accuracy" gap is a big deal for AI design.

The takeaway isn't that graphs are bad or text is bad. It's that the best explanation format depends on what kind of thinking the task requires. As AI becomes a common decision-making partner in work, health, and finance, knowing when to use text versus diagrams — and reminding users that a pretty explanation isn't necessarily a correct one — could matter as much as the AI's own accuracy.