Human-in-the-Loop Control of Objective Drift in LLM-Assisted Computer Science Education

Researchers Mark Dranias and Adam Whitley have published a paper proposing a novel curriculum to address a critical flaw in AI-assisted programming: 'objective drift.' This occurs when an LLM like GPT-4 or Claude produces locally plausible code that subtly diverges from the original task specifications, leading students astray. The paper argues that current fixes—focusing on prompt engineering—are fragile as AI tools evolve. Instead, it advocates for a permanent, human-centered approach, treating human-in-the-loop (HITL) control as a fundamental, teachable skill in computer science education.

The proposed curriculum draws from systems engineering and control theory. It explicitly separates the planning phase (where students define objectives, world models, and acceptance criteria) from the execution phase (where AI generates code). A key innovation is the intentional injection of 'concept-aligned drift' into some labs, training students to diagnose and recover from specification violations. The researchers have designed a three-arm pilot study to compare unstructured AI use, structured planning, and structured planning with injected drift, establishing detectable effect sizes for future validation. This framework aims to create durable control competencies that outlast any single AI platform.