Explainable AI is a research area that aims to make machine learning models, especially complex ones like deep neural networks, understandable to humans. The core goal is to reveal how and why a model produces a particular prediction, so that users can trust, debug, and responsibly deploy AI systems.

A key distinction is between intrinsic interpretability and post hoc explanation. Intrinsically interpretable models, such as small decision trees or linear models, are designed to be transparent from the start. Post hoc methods instead explain black box models after training, without changing their internal structure.

Popular post hoc techniques include feature importance scores that quantify how much each input feature affects the prediction, and local explanation methods that approximate the model’s behavior in the neighborhood of a single instance. Model agnostic tools that work with any classifier or regressor are widely studied, as are visualization methods that highlight influential pixels in images or attention patterns in language models.

Research also examines trade offs among accuracy, interpretability, and robustness. Making a model simpler can help humans understand it, but may reduce performance. Conversely, highly accurate models can be opaque and brittle. Recent work seeks methods that maintain high predictive power while providing faithful explanations that are not misleading.

Another important area addresses evaluation. Explanations must be judged not only by how intuitive they feel, but also by how well they reflect the true internal logic of the model and how useful they are to different kinds of users, from domain experts to laypeople. This involves human subject studies, formal metrics, and domain specific case studies.