Human-based verification strategies, such as personnel interviews and whistleblower programs, are simultaneously the most enforceable and least reliable mechanisms in this course. Interviewing staff doesn't require any additional technological infrastructure or innovation—but people can obfuscate, lie, and hide motivations. In this submodule, you will learn about the core mechanisms of human verification applicable to AI development.
Human sources are one of the collection disciplines on 2.3's map (HUMINT), and the literature files them under intelligence. They are read here as their own submodule because there is too much of them for one section of 2.3, and because audits and inspections, which take up half of this submodule, are granted access rather than intelligence: the verifier is let in, not watching from outside.
By the end of this submodule, you will be able to:
- Identify who in the AI supply chain could observe which activities, and why and how they could be incentivized to reveal, hide, or obfuscate that information.
- Trace the path information takes from reporter to verifier, and judge whether a proposed reporting channel would probabilistically motivate an insider to use it.
- Compare and contrast audits and inspections, including analyzing the knowledge boundaries of an audit or inspection regime.
- Assess the levels of independence and reliability of a verifying institution.