Talking Back -- human input and explanations to interactive AI systems

Human explanations can improve AI alignment with human values: The paper demonstrates that when humans provide explanations for their decisions to AI systems, machine learning models can learn to make judgments that better align with human concepts and reasoning, rather than optimizing solely for accuracy metrics that may diverge from human intent.

Two-way communication creates more trustworthy AI: By enabling humans to "talk back" to AI systems through explanations, the research suggests this bidirectional dialogue fosters more synergistic partnerships where AI systems become more interpretable and accountable, moving beyond one-directional explainability where AI merely explains itself to humans.

Human input shapes what AI learns to explain: The paper indicates that incorporating human explanations during model training influences not just what decisions AI makes, but also how it explains those decisions, potentially making AI explanations more meaningful and useful to human stakeholders rather than technically correct but opaque.

This approach addresses a gap in current XAI practice: While most explainable AI (XAI) research focuses on making existing AI systems more transparent to humans, this work highlights an underexplored opportunity: using human feedback and reasoning to guide AI systems toward more human-aligned decision-making from the ground up.