Researchers reveal that popular vision-language models struggle to detect misleading data visualizations, especially when deception hides in subtle caption errors.

New benchmark evaluates Vision Language Models (VLMs) on their ability to identify deceptive data visualizations paired with misleading captions

Benchmark taxonomy categorizes misleadingness into reasoning errors (cherry-picking, causal inference mistakes) and design errors (truncated axes, dual axes, inappropriate encodings)

Study combines real-world visualizations with human-curated misleading captions to enable controlled testing across different error types and deception methods

Findings suggest current commercial VLMs have significant gaps in detecting misleading visualizations, raising concerns about misinformation propagation through charts