AI Systems & Infrastructure · Axiologic Research Editions
Explainable AI
Understanding, Evaluating, and Engineering Explanations for Intelligent Systems
A guide to explanation methods, their assumptions, and the tests required before an explanation is allowed to support a consequential decision.
Accuracy is not enough
People need to know what a model is claiming, for whom, on what evidence, and with what uncertainty. A fluent rationale is not proof that a system used the reasons it presents.
Explanations as measurements
Attributions, surrogates, counterfactuals, concepts, prototypes, and generated rationales each answer different questions. The book makes their assumptions explicit and connects them to falsification, human oversight, privacy, drift, and security.
An explanation earns trust by surviving tests that could have shown it false.
From models to systems
The final sections extend the analysis to foundation models and production systems, where prompts, retrieval, tools, and interfaces all contribute to the behavior a person must understand.