Safety is the non-negotiable gate for autonomy. In 2026, regulators have moved from “let innovators test” to “prove you are safer than a human — with evidence.” Here’s the global playbook.
Global Frameworks at a Glance
| Region | Key Rule | What It Requires |
|---|---|---|
| US – NHTSA | AV STEP (2025) | Voluntary safety self-assessment + incident reporting |
| EU – UNECE | R157 ALKS & R171 DCAS | Certified automated lane keeping up to 130km/h |
| China – MIIT | ICV Pilot Cities | HD maps + V2X required for Level 4 |
| Japan | Road Transport Vehicle Act | Type approval for Level 4 (2024) |
How Safety Is Measured
Regulators now demand a safety case — not a single test. Elements:
- Miles driven in simulation (billions) + closed-track + public road
- Minimal Risk Condition (MRC): car must pull over safely if system fails
- Event Data Recorder (EDR): 30 seconds of sensor log before crash
- Independent audit by ASAM OpenX standards
Liability: Who Pays When No One Drives?
The EU's revised Product Liability Directive (2024) places liability on manufacturers for AV faults. In the US, states differ — California treats operators as “drivers.” Insurance is shifting to fleet/product policies; personal auto premiums may fall 40% as human error drops.
What Still Blocks Deployment
Uniformity. A car certified in Germany isn't auto-approved in Texas. Data sharing remains sensitive — regulators want crash logs, OEMs fear IP leak. The emerging compromise: standardized, anonymized Safety Performance Indicators (SPIs) via ISO 26262 & ISO 21448 (SOTIF).