Do Human Rationales Improve Machine Explanations?

Human input during training improves AI explanations: When humans provide explanations (rationales) to train machine learning systems, the resulting AI generates better explanations for its decisions—not just more accurate predictions. This suggests a direct link between how we teach AI and how well it can communicate its reasoning back to us.

Supervised attention outperforms unsupervised methods: The study found that text classification systems trained with human-guided "supervised attention" produced explanations that human judges rated as superior compared to systems using standard "unsupervised attention," demonstrating the measurable value of human guidance in the training process.

Bridging two separate AI fields: This research connects two previously disconnected areas—"learning with rationales" (using human explanations to improve AI performance) and "explainable AI" (making AI reasoning transparent to humans)—suggesting these approaches can reinforce each other rather than operate independently.

Human evaluation validates AI explanations: The study relied on human judges to assess explanation quality, highlighting that meaningful progress in explainable AI requires human evaluation rather than purely automated metrics, which is important for ensuring AI systems are actually understandable to real people.