At Waymo, an AI project isn’t ready until its evals are — not when the model performs well. That’s the core philosophy shared by Manasi Joshi, the company’s head of AI safety, in a recent talk at VB Transform 2026.
Beyond Benchmarks
Most AI evaluation focuses on benchmark performance: accuracy, F1 scores, human preference rankings. Waymo’s approach is fundamentally different. The company evaluates AI systems based on failure modes and edge cases rather than success rates.
“A model that achieves 99.9% accuracy but fails catastrophically in the remaining 0.1% isn’t ready for the road,” Joshi explained. “We need to understand exactly what happens in those rare scenarios.”
The Three-Layer Evaluation Framework
Waymo’s framework has three layers:
1. Simulation stress testing — millions of virtual miles testing rare scenarios: pedestrians jumping from behind buses, construction zones with contradictory signage, emergency vehicles approaching from blind spots.
2. Closed-course validation — real-world testing on Waymo’s private track with safety drivers ready to intervene. Every “disengagement” (when the safety driver takes control) is logged and analyzed.
3. Shadow mode deployment — the AI makes decisions in real vehicles but doesn’t control them. Engineers analyze what the AI would have done versus what the human driver did.
What Enterprises Can Learn
The approach has implications beyond autonomous vehicles. Any company deploying AI in high-stakes environments can adopt similar principles:
- Map failure modes first — before optimizing for accuracy, identify what could go wrong
- Test the unusual — standard benchmarks don’t capture edge cases
- Always have a human in the loop — until AI reliability reaches six-sigma levels
Waymo’s method isn’t fast or cheap. But as AI systems take on more critical tasks, the industry may need to adopt similar rigor — or face the consequences.