Assessing the Auditability of AI-integrating Systems: A Framework and Learning Analytics Case Study

Audits require systems designed for auditability from the start: The paper argues that AI systems in education need to be intentionally built to allow audits, not audited after the fact. Without this "auditability by design" approach, even well-intentioned audits cannot effectively assess whether systems are working as intended.

Three essential components enable meaningful audits: Auditable AI systems must have (1) clear, verifiable claims about how they work and their impact, (2) accessible evidence in multiple forms (documentation, raw data, system logs) to verify those claims, and (3) technical tools that allow auditors to actually access and examine this evidence.

Real-world AI systems fall short of auditability standards: When the authors assessed Moodle's student dropout prediction system, they found significant gaps: incomplete documentation about how the system works, limited ability to monitor its performance over time, and no access to test data needed to verify accuracy—problems that would prevent a thorough audit.

Lack of transparency undermines trust in educational AI: The case study demonstrates that even widely-used educational platforms integrating AI lack the documentation, monitoring capabilities, and data access necessary for independent verification, raising concerns about whether these systems are actually benefiting students as claimed.

The framework provides a practical tool for improvement: The paper offers educators and developers a concrete three-part framework to evaluate and improve auditability, supporting both assessment of existing systems and better design of future AI applications in learning environments.