This paper introduces a novel model-agnostic explanation method called anchors, which provide high-precision, local explanations for complex models. Anchors are if-then rules that represent sufficient conditions for predictions, ensuring that changes to other features do not affect the prediction. The method is designed to explain the behavior of any black-box model with high-probability guarantees. Anchors are intuitive, easy to understand, and have clear coverage, making them more effective than linear explanations or no explanations for predicting model behavior on unseen instances. The paper demonstrates the flexibility of anchors by applying them to various machine learning tasks across different domains, including text classification, structured prediction, and image classification. A user study shows that anchors enable users to predict model behavior with higher precision and less effort compared to existing methods. Anchors are also shown to be more effective than linear explanations in tasks such as visual question answering (VQA). The paper also discusses the computational efficiency of computing anchors, introducing a multi-armed bandit approach to find the best anchor with high precision and coverage. The approach is extended to a beam search method that maintains a set of candidate rules and selects the one with the highest coverage. The algorithm is evaluated on various datasets and tasks, showing that anchors provide high precision and coverage, and are more effective than linear explanations. The paper also discusses the limitations of anchors, including overly specific anchors for predictions near decision boundaries and potentially conflicting anchors. It also highlights the challenges of explaining complex output spaces and realistic perturbation distributions. The paper concludes that high precision and clear coverage are crucial for interpretable explanations of a model's local behavior, and that anchors provide a novel, effective method for achieving this.This paper introduces a novel model-agnostic explanation method called anchors, which provide high-precision, local explanations for complex models. Anchors are if-then rules that represent sufficient conditions for predictions, ensuring that changes to other features do not affect the prediction. The method is designed to explain the behavior of any black-box model with high-probability guarantees. Anchors are intuitive, easy to understand, and have clear coverage, making them more effective than linear explanations or no explanations for predicting model behavior on unseen instances. The paper demonstrates the flexibility of anchors by applying them to various machine learning tasks across different domains, including text classification, structured prediction, and image classification. A user study shows that anchors enable users to predict model behavior with higher precision and less effort compared to existing methods. Anchors are also shown to be more effective than linear explanations in tasks such as visual question answering (VQA). The paper also discusses the computational efficiency of computing anchors, introducing a multi-armed bandit approach to find the best anchor with high precision and coverage. The approach is extended to a beam search method that maintains a set of candidate rules and selects the one with the highest coverage. The algorithm is evaluated on various datasets and tasks, showing that anchors provide high precision and coverage, and are more effective than linear explanations. The paper also discusses the limitations of anchors, including overly specific anchors for predictions near decision boundaries and potentially conflicting anchors. It also highlights the challenges of explaining complex output spaces and realistic perturbation distributions. The paper concludes that high precision and clear coverage are crucial for interpretable explanations of a model's local behavior, and that anchors provide a novel, effective method for achieving this.