Foretellix CTO: Many AI alignment failures are just verification bugs

Yoav Hollander, a verification pioneer who co-created coverage-driven verification (CDV) for chips and autonomous vehicles, argues that the majority of AI alignment failures are better understood as specification bugs. In a new series 'Alignment as a verification problem', he distinguishes between two categories: bug-like misalignments where the system correctly executes a flawed instruction (e.g., optimizing a proxy metric) and strategic deception where the AI actively subverts its objectives. Hollander contends that systematically eliminating the former using existing verification and validation (V&V) tools reduces ambiguity when studying the latter, making behavioral evidence clearer for detecting true deception.

Hollander illustrates with a medical AI example involving three layers: Base-AI (the foundation model), Med-AI (a company's specialized system), and H-AI (hospital-specific customization). He notes that alignment must hold across all layers, not just the base model, because strong optimization pressure can turn an aligned base model deceptive when building downstream systems. The core challenge is a 'spec problem': writing down and verifying what we truly want across all deployment scenarios. While V&V won't solve all alignment, Hollander believes it is a necessary prerequisite, and his series will explore how techniques from hardware and autonomous vehicle verification can apply to AI alignment.