Formalizing Explainable AI: Researchers Call for Defined Problems and Metrics

Global: Formalizing Explainable AI: Researchers Call for Defined Problems and Metrics

A team of researchers led by Stefan Haufe has highlighted fundamental shortcomings in the current practice of explainable artificial intelligence (XAI) and urged the community to adopt a more formalized approach. Their analysis, presented in a pre‑print on arXiv, argues that many popular XAI techniques fail to answer the most relevant questions about machine‑learning models, their training data, and test inputs.

Scope and Goals of XAI

Explainable AI seeks to make the decisions of complex machine‑learning systems understandable to human users, thereby supporting trust, accountability, and insight. The authors note that while the field has grown rapidly, its methods often lack clear problem definitions and rigorous evaluation standards.

Systematic Attribution Issues

The paper points out that existing XAI methods frequently assign importance to input features that are statistically independent of the prediction target. This systematic misattribution limits the ability of explanations to diagnose model behavior, correct data issues, or uncover scientific insights.

Consequences for Practice

Because of these limitations, practitioners may find XAI outputs insufficient for tasks such as model debugging, scientific discovery, or identifying actionable intervention points. The authors contend that without addressing these gaps, the utility of XAI remains constrained.

Call for Formal Problem Definition

To remedy the situation, the researchers propose that developers explicitly define the specific explanatory problem they aim to solve. They suggest that different use cases will require distinct notions of explanation correctness, accompanied by objective performance metrics.

Proposed Evaluation Framework

According to the authors, future XAI work should incorporate targeted criteria that assess whether explanations meet the defined correctness standards. Such metrics would enable systematic validation of XAI algorithms across diverse applications.

Implications for Future Research

If adopted, this formalization could steer the development of more reliable and purpose‑aligned explanation methods, fostering greater confidence in AI systems across scientific, industrial, and regulatory contexts.

Publication Details

The manuscript was first submitted on 22 September 2024 and underwent revisions, with the latest version posted on 9 January 2026. It is classified under Machine Learning and Artificial Intelligence on arXiv.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.