Towards Human-Understandable Visual Explanations:Imperceptible High-frequency Cues Can Better Be Removed

Current AI explanations may not actually be understandable to humans: The paper finds that neural networks often rely on "distinguishing features" that are invisible to the human eye—specifically high-frequency visual cues—to make predictions, even though standard explainability methods highlight these features as important. This means existing AI explanations can appear to explain a model's reasoning while actually describing something humans cannot perceive.

Human visual limitations should be built into AI systems from the start: Rather than explaining what a trained AI learned after the fact, the researchers propose training neural networks to focus on features that humans can actually see and understand. This "human perceptibility principle" suggests that responsible AI development requires considering human perception capabilities during model design, not just during explanation.

Human-centered AI design can increase trust and enable mutual learning: In a user study on detecting fake faces, the human-perceptible model "better aligned with human intuition" compared to standard models. This suggests that when AI systems are designed with human perception in mind, people may find them more trustworthy and be better able to learn from the AI's reasoning—potentially creating a more collaborative relationship between humans and machines.

Imperceptible high-frequency patterns are a hidden problem in current AI: The paper specifically identifies that neural networks exploit subtle, high-frequency visual patterns that humans cannot consciously detect. Making this invisible reliance explicit is important for understanding where current AI explanations fall short and why some "explained" AI decisions may still feel opaque to users.