This research introduces the first benchmark for evaluating how multimodal large language models detect and interpret misleading chart visualizations that can distort perceptions.

• Establishes a systematic framework to test AI models' ability to identify deceptive visual tactics
• Evaluates leading MLLMs on their chart comprehension capabilities
• Reveals current limitations in AI systems for detecting visual misinformation
• Provides groundwork for developing more robust visual information security tools

Security Impact: As misleading data visualizations become more sophisticated, this benchmark helps develop systems that can automatically detect visual misinformation, protecting information integrity and reducing the spread of deceptive analytics in business and public discourse.

Unmasking Deceptive Visuals: Benchmarking Multimodal Large Language Models on Misleading Chart Question Answering