On July 30, 2026, a document designated GB 44721-2026 was officially approved.
This string of numbers may look unremarkable, but for China's entire automotive industry, it signifies one thing: the era of "wild, unchecked growth" for autonomous driving has come to an end.
From "Recommended" to "Mandatory": A World of Difference
Let's first look at the background of this standard.
In 2024, the industry already had a national standard related to autonomous driving, designated GB/T 44721-2024. Note the "/T" in the middle-this indicates a "recommended" standard. Put simply, a recommended standard suggests a course of action, but there are no consequences if you choose not to follow it. Automakers could choose whether or not to reference it; it carried no legal binding force.
What is the current level of autonomous driving technology? – Sina Auto
However, the newly released GB 44721-2026 omits that "/T." The shift from "recommended" to "mandatory" represents a world of difference.
Starting July 1, 2027, any new vehicle equipped with Level 3 (conditional automation) or Level 4 (high automation) systems must meet all the requirements of this standard to be sold in the Chinese market. Fail to meet the standard? Sorry-production and sales are prohibited, and non-compliance constitutes a violation of the law.
This is not a suggestion; it is a bottom-line requirement.
What exactly do L3 and L4 mean?
Before diving into the details of the standard, we need to clarify these two levels.
According to the internationally accepted classification of autonomous driving, Levels L0 through L2 are classified as driver assistance-the human drives, while the system provides help. L3 is "conditional automation"-the system can take full control of driving under specific conditions, but the driver must be ready to take over at any moment. L4 is "high automation"-the system can independently perform all driving tasks in specific scenarios without requiring the driver to constantly monitor the situation.
What are the levels of driver assistance? – Yiche Encyclopedia
To use an analogy: L3 is like a co-driver acting as an instructor; they do most of the driving, but if they encounter a situation they can't handle, they'll call on you to take the wheel. L4 is like a dedicated chauffeur; on designated routes, you can essentially leave everything to the system. The vast majority of vehicles currently on the market marketed as "autonomous" are actually only at Level 2 (L2) autonomy. True Level 3 (L3) and Level 4 (L4) technologies have previously existed in a state of "having the tech but lacking the rules" in China; automakers were capable of building them but hesitated to sell them at scale because there were no clear guidelines on liability in the event of an accident.
**Four Major Safety Requirements: Non-Negotiable Benchmarks**
The new standard establishes a comprehensive "safety assessment," comprising four dimensions, each with strict, non-negotiable requirements.
**First, automakers must manage safety from the very foundation.**
The standard requires manufacturers to build safety assurance capabilities covering the entire lifecycle-from product design and manufacturing to post-deployment-across four dimensions: safety policy, risk management, safety assurance, and safety improvement. In plain language: automakers cannot simply boast at product launches that "our autonomous driving is safe"; they must demonstrate a complete safety management system spanning from R&D to after-sales service, controlling risks right from the source.
**Second, autonomous driving must not perform worse than a human driver.**
The standard explicitly requires that the safety level of an autonomous driving system must at least match that of a competent and attentive human driver. Furthermore, the system must not pose unreasonable safety risks to other road users.
**Third, L3 systems must include a "driver monitoring" function.**
This is the point most directly relevant to the average consumer: Level 3 autonomous driving systems must be equipped with a function to monitor the driver's ability to take over control.
Put simply, the car has to keep an eye on you. While the system is operating, it must ensure you are ready to take the wheel at any moment. If you fall asleep, play on your phone, or become distracted, the car must detect this and alert you. This isn't about spying on you; it ensures that if the system encounters a situation it cannot handle, you can take over in time. At the same time, automakers must clearly inform users-via owner's manuals, official websites, and in-vehicle displays-about what the vehicle's autonomous driving system can and cannot do, and under what circumstances the driver must take control.
Fourth, new vehicles must pass "three hurdles" before hitting the market.
The standard establishes a tripartite inspection system comprising "assessment of corporate assurance capabilities," "safety dossier review," and "confirmatory testing."
[Image]
Of these, the "safety dossier" warrants particular attention. The new standard includes a dedicated 17-page appendix specifying how these dossiers must be compiled. Automakers are required to use a "declaration-argument-evidence" structure to clearly articulate system functions, associated risks, verification methods, and post-deployment monitoring and improvement strategies.
A dossier alone is insufficient; third-party testing agencies must also conduct proving-ground tests, on-road trials, and simulations to comprehensively evaluate the system's actual performance.
A Paradigm Shift for the Industry
The introduction of this standard compels the entire industry to adjust.
In recent years, the focus was on accumulating autonomous driving mileage, showcasing flashy features, and being the first to achieve specific milestones. Such "mileage races" will no longer suffice; the primary battleground of industry competition is shifting from a "performance race" to a "safety race."
Experts predict that non-compliant companies will be weeded out, exaggerated marketing claims will cease, and the supply chain will undergo a major shake-up.
Compliance costs are also skyrocketing. With the transition from a recommended standard to a mandatory one, the compliance cost per vehicle starts in the tens of millions of yuan. For major manufacturers already invested in advanced intelligent driving, this presents an opportunity to solidify their competitive edge; for companies lacking sufficient technical depth, however, these compliance expenses could prove to be the "last straw."
From another perspective, however, this is good news for consumers. Previously, purchasing a vehicle with advanced intelligent driving capabilities was akin to buying a "blind box"-one could never be certain of the system's reliability. With the new mandatory national standard in place, every L3 or L4 vehicle on the road will have passed a standardized safety assessment.
July 2027: The Curtain Rises
The standard grants the industry a grace period of approximately one year. Coincidentally, major automakers-including Huawei-affiliated brands, Xpeng, GAC, Zeekr, and BYD-have all designated the 2025–2027 period as the critical window for the mass production of L3 and L4 autonomous driving technologies. GAC has even publicly announced plans to launch an L4 model for private consumers by 2027.
The timing is impeccable.
After July 1, 2027, L3 and L4 autonomous driving in China will move beyond the stages of mere "testing" and "piloting," becoming subject to unified legal standards.
Consider this question:
When a car drives itself, how is liability apportioned in the event of an accident? The answer to this question is gradually being clarified by these standards. By requiring automakers to establish safety assurance capabilities covering the vehicle's entire lifecycle, the standards essentially reinforce the automakers' primary responsibility. This signals the end of the old paradigm where the "driver bears full responsibility" in the era of L3 and L4 autonomy.
However, how will rights and responsibilities be defined for specific accidents? Will automakers' safety records withstand scrutiny? Can data monitoring regarding takeover capabilities serve as a basis for determining liability?
These questions remain unanswered.





