Human-in-the-loop model explanation via verbatim boundary identification in generated neighborhoods

Human oversight improves AI explanations: Rather than relying solely on automated explanation methods that may be inaccurate, this approach incorporates human judgment at three intervention points, allowing people to actively refine and explore why a model made a specific decision.

Local decision boundaries reveal model behavior: By generating synthetic examples around a given input and observing how the model classifies them, the method maps out the "local decision boundary"—essentially showing what small changes would flip the model's prediction, providing more faithful insights than approximate explanations alone.

Addresses the black-box problem in high-stakes applications: The paper identifies that opaque machine learning models create "trust crises" in critical domains, and this human-in-the-loop approach aims to make model decisions more transparent and interpretable for case-critical applications where understanding why a decision was made is essential.

Synthetic examples make explanations concrete: By generating plausible neighboring instances and showing how the model responds to them, the approach provides "multifaceted understanding" of model behavior in a way that is more tangible and verifiable than abstract feature importance scores or other traditional XAI methods.