Machine Unlearning Testing Framed as Software Engineering Challenge

Machine learning components have become fundamental to AI-driven software systems, including recommendation engines, coding assistants, and clinical support tools. Regulatory requirements and governance frameworks increasingly demand the removal of sensitive data from deployed models, making machine unlearning a practical alternative to complete retraining. This approach presents a significant software quality-assurance challenge: under real-world deployment constraints with imperfect oracles, how can we verify that a model no longer depends on targeted information? The paper positions unlearning testing as a primary software engineering concern. Practical unlearning tests must offer comprehensive coverage across proxy and mediated influence pathways. They should provide debuggable diagnostics that identify where data leakage persists. These tests need to be cost-effective for regression-style execution within query budgets. Additionally, they must be applicable in black-box scenarios for API-deployed models. The research outlines a causal framework to address these requirements.