NHTSA has expanded its investigation into Tesla's Full Self-Driving system to approximately 3.2 million vehicles. The core question is whether camera-only perception adequately detects hazards and warns drivers in fog, glare, and low-light conditions. This is not a routine recall inquiry. The agency is treating poor-visibility behavior as a systemic fleet characteristic rather than an isolated defect.
Tesla's long-standing defense has been cumulative deployment volume. The company points to roughly ten billion miles of FSD data as evidence of safety equivalence—or superiority—over human drivers. NHTSA's widening scope suggests that argument is losing its persuasive force. When regulators ask whether cameras can see through fog, "we have lots of data" is not an answer. It is a category error.
The shift here is from breadth-based to depth-based validation. Aggregate incident rates once served as a statistical shield. Now the agency demands demonstrated performance in specific environmental edge cases. The probe asks whether depth of validation can be inferred from breadth of deployment, and NHTSA's answer appears to be negative.
Waymo offers the regulatory counterexample. California approved its expansion across eighteen counties this week, rewarding a tightly bounded operational design domain backed by lidar, radar, and third-party safety submissions. Rather than converging, the two approaches are being codified into divergent regulatory tiers. One path rewards demonstrated competence in well-defined conditions; the other faces scrutiny for claiming generality without proving it.
The ten billion mile figure deserves scrutiny. It measures exposure, not understanding. A system can drive many miles while failing to recognize the same hazard repeatedly, if the failure mode is rare enough to avoid statistical notice. NHTSA seems to recognize this. The investigation targets the gap between statistical safety and functional reliability—the difference between "usually works" and "works under stated conditions."
For autonomous vehicle policy, this matters. If validation requires depth, not just breadth, the economics of deployment change. Demonstrating competence in specific conditions is more expensive than collecting aggregate miles. It requires structured testing, defined operational domains, and third-party verification. The regulatory framework is bifurcating, and the path forward favors those who can prove their boundaries rather than those who claim to have none.
Sources