Explainable Active Learning (XAL): An Empirical Study of How Local Explanations Impact Annotator Experience

Explanations improve trust calibration: The study found that when machine teachers received explanations of the model's reasoning alongside predictions, they developed more accurate trust in the system—neither over-trusting nor under-trusting it—compared to traditional active learning without explanations.

Human feedback quality changes with explanations: The research demonstrated that the content and nature of feedback provided by annotators differed when they could see the model's reasoning, suggesting that explanations fundamentally alter how humans engage in the teaching process.

Explanations address a critical gap in human-AI interfaces: The paper identifies that current active learning systems have "minimal and opaque" human-AI interfaces, and proposes that explainable AI techniques can serve as a core solution to make the machine learning training process more transparent and collaborative.

The "machine teacher" role requires understanding model reasoning: Just as human students benefit from teachers explaining their reasoning, the study argues that human annotators training ML models need to understand why models make mistakes to effectively guide their learning and feel confident in their teaching outcomes.

Explainable active learning outperforms standard approaches: The empirical comparison showed that XAL produced better model learning outcomes and annotator experience than both traditional active learning and systems that only showed predictions without explanations.