Explaining Explaining

Current AI systems cannot truly explain their decisions: Machine learning models that power most AI today are "black boxes" that lack the ability to provide genuine explanations for their outputs, which is particularly problematic in high-stakes domains where people need to understand why decisions are made.

Redefining "explanation" creates a false solution: The explainable AI (XAI) movement addresses the black box problem by changing what "explanation" means rather than solving the underlying issue, potentially giving users false confidence in systems that still cannot truly justify their reasoning.

Human-centered AI has good intentions but limited reach: While the human-centered explainable AI (HCXAI) movement correctly identifies what users actually need from explanations, it remains constrained by its commitment to machine learning and therefore cannot fully deliver on those needs.

A hybrid approach combining knowledge and data offers a path forward: The paper proposes using knowledge-based systems (which can explain their logic) supplemented by machine learning where appropriate, rather than relying on machine learning alone, to create AI that can provide meaningful explanations to users.

Humans must retain ultimate responsibility: In this hybrid model, AI systems serve as assistants to human decision-makers who bear final responsibility for outcomes, rather than autonomous systems making critical decisions independently.