"接近十倍" — nearly ten times. That's reportedly Peng Jun's phrase, reportedly delivered from a conference stage in Shenzhen around May 29, describing how much lower Pony.ai's accident rate runs versus human drivers.
It is not a filing. It is not audited. It is reportedly a founder's line at the Future Automotive Pioneers Conference, and this week the market treated it as a data point.
Two China headlines landed together and read like a matched set. Beijing sent a national self-driving liability framework to the Standing Committee of the National People's Congress on Tuesday. Pony.ai reported robotaxi revenue up 691.2 percent year-on-year to US$12.1 million.
The tidy story writes itself: the law legitimizes the technology, the growth proves the market, and China's autonomous-vehicle moment has arrived.
Last week's open thread here was which regulator would be first to permit an empty driver's seat at national scale. China's draft is that framework arriving. But the answer is still incomplete, and the reason both headlines share the same crack.
Both rest on an assumption neither one proves: that someone, after the fact, can independently verify what the car was actually doing when something went wrong.
Start with the law, because on paper it does the hard thing. The draft adds a dedicated chapter—nine chapters, 170 articles total—and says that when a traffic violation occurs while a fully autonomous function is activated, the manufacturer or importer handles it. If the automaker claims the violation was unrelated to the autonomous system, the burden is on them to prove it.
That is a real shift. It moves first-line responsibility from the human in the seat to the company that built the system, exactly as Peng Jun himself framed the core distinction of driverless vehicles: the first responsible party changes from the driver to the car.
Then comes the sentence that quietly hollows out the framework. As CnEVPost notes from the public summary, the draft "does not explain how vehicle operating data would be accessed or what evidentiary rules would apply in disputes."
Read that against the burden-shifting clause and the problem is obvious. The whole liability regime hinges on one binary fact—was the autonomous mode active or not—and the draft is silent on who gets to see the data that answers it.
The automaker holds the logs. The automaker also carries the liability. And the automaker is the party arguing, in any dispute, that its system wasn't at fault. Nothing in the public summary breaks that circle.
China Daily flags why this matters beyond individual crashes. The draft also bars manufacturers from "making false or exaggerated claims about autonomous driving functions." Cui Dongshu, of the China Automobile Dealers Association, notes that terms like "intelligent driving" and "automatic parking" have blurred what systems actually do.
So the law creates two obligations it cannot yet enforce: prove which mode was active, and don't overstate what the system can do. Both require evidence rules that don't exist on the page.
The whole liability regime hinges on one binary fact—was the autonomous mode active or not—and the draft is silent on who gets to see the data that answers it.
Now put Pony.ai's numbers next to that void, because they occupy exactly the same unverified space.
The 691.2 percent figure is real and the underlying growth is real—fare-charging revenue rose 849.3 percent. But $12.1 million of robotaxi revenue sits inside a total net loss of $45.4 million for the quarter. The loss narrowed from $53.3 million a year earlier; it did not disappear.
The fleet that generated the headline stood at 1,975 vehicles as of June 30, targeting more than 3,500 by year-end. In Shenzhen, Pony.ai reportedly reported single-vehicle averages of about 25 orders per day in dense areas such as the airport, Shenzhen Bay port and Shekou port.
That is a small base carrying a very large claim. "Nearly ten times lower" than human accident rates, off a fleet under two thousand vehicles, is a statement that would require enormous exposure and rigorous methodology to defend—and it was delivered as a speech, not a disclosure.
No regulator has yet defined how anyone would check it. The MIIT standard taking effect July 1, 2027, requires autonomous systems to be "at least equivalent to a qualified and attentive driver"—a benchmark, not an audit protocol. It sets the bar without specifying who measures the jump.
This is where the two stories stop being mutual validation and start being the same unanswered question wearing two costumes.
The law says the automaker must prove its system wasn't involved, but doesn't say how the data gets pried loose. The revenue story says the fleet is nearly ten times safer than humans, but the number that would prove it lives entirely inside the company.
In both cases the party with the strongest interest in the answer is the only party holding the evidence.
The business model sharpens the point. Pony.ai's surge is explicitly framed as an asset-light push—offloading vehicle capital through partners like the Uber deal covering more than 2,000 robotaxis in Europe and the ComfortDelGro launch in Singapore, with a pending pipeline over 4,000 vehicles.
Asset-light means the capital risk moves outward to partners while the safety claims stay in-house. Growth accelerates precisely because someone else owns the cars—but the accident data, the operating logs, the methodology behind "ten times" never leave Pony.ai's servers.
That is the structural mirror of the law's gap. Distribute the vehicles, centralize the evidence.
None of this argues the technology doesn't work or that the growth is fake. It argues that the two loudest signals of the week—a liability framework and a revenue multiple—both certify momentum, not safety.
The draft is genuinely a milestone. It drags China, per Reuters, from a patchwork of Shenzhen and Beijing local rules toward a unified national regime, and analyst Gao Shen reportedly expects "dozens of new autonomous models" once it's promulgated. Commercial clarity is arriving.
But clarity about who handles a violation is not the same as clarity about proving the violation happened in autonomous mode. The draft solves the assignment problem and leaves the evidentiary problem for what CnEVPost calls "intense negotiation between automakers, insurers and regulators."
Until that negotiation produces access rules and evidentiary standards, the accountability chain has a missing link right where it matters most.
So here is the real test the law sets without meeting. The first genuine milestone in Chinese AV commercialization won't be a revenue multiple, however large.
It will be the first published, third-party-auditable safety case—a "ten times safer" claim that a regulator or independent body can actually verify against accessible operating data. Or it will be the first enforcement action under that new ban on "false or exaggerated claims," which would force the question of what evidence proves a claim false.
Whichever comes first, that's the moment the trust problem gets solved—not the moment a founder says "接近十倍" from a stage, and not the moment a revenue line grows 691 percent off a fleet of 1,975 cars.
The framework has arrived. The verification has not. And you cannot legislate accountability for a mode you have not yet decided how to prove was on.