China's Ministry of Industry and Information Technology (MIIT) and the State Administration for Market Regulation (SAMR) approved GB 44721-2026, the country's first mandatory national safety standard for L3 and L4 autonomous driving systems, on July 30, 2026. It takes effect July 1, 2027, and applies to M-class (passenger) and N-class (goods) vehicles running conditional or high driving automation, excluding automated parking systems.
What the standard actually requires
The rule sets a performance floor: an autonomous system must be "at least as safe as a competent and attentive driver performing the dynamic driving task" and cannot introduce unreasonable risk. Manufacturers must build safety management across the full product lifecycle — design, production, and post-deployment monitoring — organized around what regulators call safety policy, risk management, safety assurance, and safety improvement.
Three provisions matter most for how liability and trust get established after something goes wrong. First, vehicles must run a Data Storage System for Automated Driving (DSSAD) — a flight-recorder-style black box logging sensor inputs, system status, and control handoffs, building on a separate mandatory data-recording standard that took effect in January 2026. Second, L3 systems require active monitoring of driver readiness to take over, with defined activation and deactivation states rather than an ambiguous handoff. Third, manufacturers must disclose automation levels, capabilities, and limitations to buyers through manuals, in-vehicle displays, and official channels — an explicit rejection of marketing language that blurs assisted driving with full autonomy. Compliance will run through a mixed testing framework: corporate safety-capability checks, safety-dossier review, and confirmatory road, simulation, and field testing (miit.gov.cn, std.samr.gov.cn).
The crash that forced the question
The standard did not emerge in a vacuum. On March 29, 2025, a Xiaomi SU7 traveling at 116 km/h on the Dezhou–Shangrao highway in Anhui province was operating in Navigate on Autopilot mode when it approached a closed construction zone. The system issued a hands-off warning at 10:36:48 p.m. and a risk warning for the obstacle ahead at 10:44:24 p.m. — one second before the driver took manual control at 10:44:25 p.m. Despite braking and steering, the car struck a concrete barrier at an estimated 97 km/h and caught fire, killing three university students (carnewschina.com). That one-second gap between warning and takeover is precisely the failure mode the new standard's driver-readiness-monitoring requirement targets. Xiaomi went on to recall roughly 116,900 SU7 vehicles in September 2025 after SAMR found the highway assisted-driving system showed "insufficient recognition, warning and handling ability" in extreme conditions (scmp.com).
The case for the rule
The strongest argument for GB 44721-2026 is straightforward: without a black box, there is no reliable way to establish whether the car or the driver was in control at the moment of a crash, and without a real safety floor, marketing claims about "intelligent driving" can outrun what the hardware actually does. China's own regulators already banned automakers from using the terms "intelligent driving" and "autonomous driving" in advertising after the SU7 crash. A verifiable data trail protects consumers, gives investigators and courts something to work with, and — done right — should make it easier, not harder, for capable systems to prove themselves against a clear, uniform bar rather than a patchwork of provincial pilot rules.
Where the rule risks overreach
The risk is less in the black box than in the compliance architecture around it. A single, centralized "China-specific test scenario library" and mandatory third-party confirmatory testing, layered on top of the safety-dossier review, could become a bottleneck that favors incumbents with the regulatory-affairs staff to navigate it and disadvantages smaller or foreign entrants who must now certify against a bespoke domestic framework rather than reference an international one such as UN R157. A one-year runway from approval to enforcement — July 2026 to July 2027 — is tight for manufacturers who need to redesign data architecture, not just firmware, across model lines already in production.
The proportionate path
The DSSAD and disclosure requirements are the right instrument for the right problem: they target information asymmetry between manufacturers and both regulators and consumers without prescribing how the underlying autonomy stack must work. That is proportionate regulation — it addresses the actual failure in the Xiaomi case, an opaque handoff moment, without dictating sensor suites or software architecture. The testing and certification apparatus is the part worth watching. If SAMR keeps the confirmatory-testing bar focused on the safety floor already written into the standard rather than expanding it into a de facto localization requirement, China will have produced a genuinely exportable model: mandate the evidence, not the engineering. If the testing regime instead becomes a market-access lever, the safety rationale will have been real, but the effect on competition will not have been incidental.