AI Audit Lab: A Phase 1 Design-and-Development Framework and Evaluation Protocol for Auditing AI-Assisted Machine-Learning Workflows

A Phase 1 Design-and-Development Framework and Evaluation Protocol for Auditing AI-Assisted Machine-Learning Workflows

Abstract

Generative AI (GenAI) coding assistants let learners produce working machine-learning notebooks in seconds, but the generated code can contain subtle methodological errors — data leakage, invalid validation splits, unwarranted causal claims — that read as entirely convincing. Left unaddressed, uncritical reliance on GenAI output risks normalizing practices that have already produced overoptimistic, irreproducible findings across dozens of scientific fields (Kapoor & Narayanan, 2023). This paper reports the Phase 1 design and development, following Richey and Klein (2007), of the AI Audit Lab: an instructional framework that repositions GenAI output from an answer to be trusted into an artifact to be audited. Learners work through a five-step cycle — inspect, predict, test, repair, explain — using notebooks seeded with realistic, GenAI-style errors, supported by a Socratic AI tutor that withholds the fix until a hypothesis is committed. The framework specifies eight error scenarios across three validity-failure categories, paired with feedback tiered by evidentiary strength. We describe the system architecture; report a seeded, computationally reproducible demonstration that two leakage types produce very different numerical consequences despite both being invalid; and specify a within-subject Phase 2 evaluation protocol, including a six-threat model, for a future comparison against standard AI tutoring. This paper makes no claim of pedagogical effectiveness; its materials are intended as an openly reusable resource once the Phase 2 evaluation has been conducted.