The integration of large language models into learning-intensive domains has created a fundamental measurement problem: proxy failure. A polished artifact—whether coursework, research output, or professional work—may demonstrate technical competence in leveraging AI while providing zero credible evidence of the underlying human understanding, judgment, or transfer capacity that such work purportedly certifies. Current governance frameworks lack the granularity to address this distinction, treating AI-assisted outputs monolithically rather than decomposing them into artifact residual (the final deliverable's intrinsic quality) and capability residual (demonstrable human cognition).
AI to Learn 2.0 operationalizes this distinction through a five-component deliverable package architecture: usability (functional independence from the original generative model), auditability (transparent reasoning chains reconstructible without cloud APIs), transferability (applicability to novel problem instances), justifiability (human-articulated rationales for design choices), and context-appropriate evidence (explanation depth or transfer demonstrations calibrated to domain requirements). The framework implements these constraints via a seven-dimensional maturity rubric with gating thresholds on critical dimensions—preventing progression to higher certification levels until foundational capability evidence accumulates.
The approach deliberately permits opaque AI during exploratory phases (hypothesis generation, drafting, workflow design) while imposing deterministic quality gates on released deliverables. This asymmetry reflects the pedagogical reality that scaffolding and iteration benefit from generative assistance, but final outputs must demonstrate auditable reasoning. The accompanying capability-evidence ladder operationalizes this by mapping domain-specific tasks to appropriate evidence types—ranging from symbolic-regression problems requiring mathematical justification to lecture-to-quiz pipelines demanding reproducible assessment protocols.
Validation across contrastive cases—including coursework substitution scenarios and teacher-audited national examination materials—demonstrates how the framework systematically separates substitution workflows (high polish, low capability residual) from bounded, handoff-ready AI-assisted work. This positions AI to Learn 2.0 as a structured instrument for third-party review where validity preservation and accountability boundaries are non-negotiable.
