You have learned the main methods of human-based AI verification; traced the reporting channel from whistleblower to verifier; and distinguished where, when, and what types of inspections are effective. In 2.4.4, the last section of this submodule, you will put together everything you've learned to form judgments about auditing and reporting institutions as a whole. The question is: can the inspecting institution reach a defensible and reliable judgment about the evidence it handles?
Explore the below frontier AI auditing design principles and a historical nuclear parallel for features of effective institutions.
Frontier AI auditing: design principles
Read the eight design principles in the Executive Summary, from Scope of Risks through Clarity.
Brundage et al. (2026) | 6–8 min
Defining Noncompliance: NPT Safeguards Agreements
Read “Determining noncompliance” and “Standard of proof”, and keep the five-case table beside them.
John Carlson, Australian Safeguards and Non-Proliferation Office (2009) | 6–8 min