Once per 160 kilometers. That is the intervention rate South Korea just wrote into law — a regulator putting an actual number on paper for what "driverless" has to mean before a Level 4 vehicle can carry the public without a driver on board.
Hold that number in mind, because it is the rarest thing in this industry: a verifiable threshold. Almost nowhere else does one exist.
Last week, we mapped the fragmentation of global AV rules — every jurisdiction writing its own rulebook, no shared definition of "safe enough." This week the fracture turned into a two-continent fight, with Tesla at the center of both fronts.
Same company. Same week. Expansion and doubt arriving on the same page — the fracture in global AV rules turned into a two-continent fight.
Start with the calendar. In a single seven-day stretch in early July, Tesla put monitor-free robotaxis on the streets of Miami's western fringe — no safety operator in the car — while US senators demanded it verify the FSD safety numbers it advertises, and EU regulators were told those same numbers had been inflated to win European approval.
Same company. Same week. Expansion and doubt arriving on the same page.
The Miami launch followed Tesla's now-familiar pattern: avoid the busy downtown, start with a small zone, no in-car monitor. Riders posted that they'd taken unsupervised trips. It reads like confidence.
But the fleet math undercuts the swagger. In Texas, Tesla holds authorization for just 69 unsupervised robotaxis — and reports suggest the active count has been drifting down, not up. The expansion is geographic, not volumetric. New cities, thin fleets.
Now set that against what Waymo published the same week — because the contrast is the whole story.
Waymo released two studies, peer-reviewed and accepted in Traffic Injury Prevention, built on a single deceptively simple premise: not all miles are equal. A highway commute on Tuesday morning is not a downtown fare at 2 AM on a Saturday, and comparing an AV's blanket average to a human's blanket average tells you nothing.
So Waymo paired human crash databases with granular traffic-volume data to reconstruct exactly when and where humans actually drive, then matched its own record against that.
The findings are the point. On surface streets, human drivers in Memphis were involved in fatal crashes at 8.4 times the rate of drivers in Boston — meaning a national average overstates Boston's risk threefold and understates Memphis's by the same margin.
Time matters just as much. Between midnight and 4 AM, human crash rates surge 2 to 5 times higher on weekdays and 2.5 to 6 times higher on weekends versus the daily average — a window invisible in aggregate data because overnight driving is only 1.5% of total human mileage.
That masking is the sleight of hand Waymo is calling out. And it cuts directly at Waymo's own position: its fleet drives proportionally four times more overnight miles than the average human, placing it disproportionately in the most dangerous hours.
Across 127 million autonomous miles, compared against humans navigating the same locations, days, and times, Waymo says it still came out ahead in every time window analyzed.
Whether or not you take that number at face value, notice what Waymo actually did: it defined the benchmark, published the method, and invited scrutiny. It raised the evidentiary bar for what a credible safety claim even looks like.
Tesla is on the other pole. Its numbers aren't being published to a bar like that — they're being contested on two continents at once.
In Washington, senators want to check whether Tesla's FSD safety marketing reflects reality or misleads the public. In Europe, Tesla stands accused of submitting inflated safety data to secure FSD approval. Same claims, two regulatory bodies, one credibility problem.
The pattern is stark. Waymo scales by disclosing more. Tesla is scaling fastest — Austin, Dallas, Houston, now Miami — while disclosing least, and its available disclosures are precisely the ones under subpoena and challenge.
That's the uncomfortable engine of the 2026 expansion. Permission is running ahead of verification.
Look at what regulators are loosening. In the US, the government is moving to drop the requirement that autonomous cars carry a brake pedal — clearing a physical obstacle to purpose-built driverless vehicles, and doing it on the assumption that the safety data flowing in is honest.
The competitive field is crowding the same open door. Baidu's Apollo Go is preparing to launch in Switzerland and Turkey, eyeing Europe's high driver-labor costs as the thing that finally makes robotaxi economics work. Zoox is running its fleet in Las Vegas and testing toward Miami. Waymo has been in Miami since January.
Everyone is expanding into whichever jurisdiction opens first. Which is exactly why the definition of "open" matters.
South Korea is the outlier that exposes the rest. Its new Ministry of Land guidelines don't gesture at safety — they quantify it: a 15,000-kilometer minimum log, an intervention rate no worse than once per 160 km, real-time remote monitoring, and a defined minimum-risk-condition strategy for bringing a stalled vehicle safely to a stop.
The rules even show their work on the leniency. The 15,000 km can be pooled across up to five vehicles running identical systems, each logging at least 3,000 km. That's a regulator making a deliberate trade-off in the open, with the numbers visible.
Compare that to the American approach — dropping a hardware mandate while the leading disclosed dataset from Tesla is simultaneously accused of being inflated. One regulator is writing thresholds; the other is removing them.
But even South Korea's number exposes the deeper hole. One intervention per 160 km is measurable only if everyone agrees on what counts as an intervention, who logs it, and who audits the log. Korea drew the number from overseas Level 4 permit regimes — but "intervention" has no shared, audited definition across vendors or borders.
So the entire edifice rests on a metric no one has standardized and no one independently verifies. The rules are being rewritten around a figure the industry hasn't agreed how to define.
This is the same pathology now visible in the humanoid sector, where companies quote reliability and uptime numbers no one will define or let anyone check — undefined safety metrics driving capital and permission in lockstep.
For robotaxis the stakes are more immediate, because the vehicle is already carrying a paying passenger through downtown Miami with no one in the driver's seat.
Here is where the two poles resolve into a single verdict. Waymo's method is falsifiable — it published the benchmark and the miles, so someone can check it. Tesla's is not yet, because the numbers are contested precisely where they'd need to be trusted.
Until a Tesla-grade dataset survives the kind of scrutiny Waymo's own methodology invites — time-matched, geography-matched, peer-reviewed, auditable — "driverless" remains a permitting status, not a verified safety claim.
The Miami cars are real. The 69 Texas authorizations are real. The subpoenas and the EU inflation charge are real.
What's missing is the one thing regulators on two continents are loosening around: a number they can actually check. Expansion is a decision. Safety is a measurement. In 2026, the industry is racing on the first and hoping no one asks too hard about the second.