Here is the contradiction most coverage of the AV Emergency Response Coordination Act glides past: everyone defending robotaxis reaches for the same statistic, and it measures the exact thing that isn't broken.
According to a study of operations between 2021 and 2024, Waymo's cars reportedly crash 68% less often than human drivers across 50 million miles. That number is real, and it is beside the point. Nobody proposing this bill is worried about the crashes. They are worried about the stalls.
Watch the two facts sit next to each other and the seam opens immediately. The industry answers "why should we trust these cars?" with collision data, while the failures piling up in San Francisco and Wuhan have nothing to do with collisions and everything to do with what a car does when its brain freezes.
The industry answers crashes with collision data, while the failures piling up in San Francisco and Wuhan have everything to do with what a car does when its brain freezes.
Last week, this thread was a regulatory posture: NHTSA had told Waymo, Zoox, and others their interference with first responders was "unacceptable," demanded solutions by month's end, and the deadline passed. This week there is a named federal bill built specifically for the emergency scene. The delta matters. The question is whether the bill closes the gap or just formalizes it.
Start with the tell.
On the steps of Fire Station 4 in Mission Bay, Mayor Daniel Lurie was delivering the industry's careful both-sides line. "We have seen the enormous potential of this technology," he said, "we have also learned where stronger coordination and clearer standards are needed."
Shortly after, per the Chronicle, a Waymo drove by and briefly blocked a line of cars trying to pass. The failure mode drove into the frame moments before he finished, on cue.
Here is what Rep. Kevin Mullin's bill actually does. It requires AV companies to publish emergency response guides telling first responders how to contact the company, how to move a stalled vehicle, and how to reach someone trapped inside. It mandates a 24/7 hotline for direct contact with operators. It creates digital geofencing so that, reportedly, officials can request that robotaxis avoid active emergency zones. And it directs NHTSA to set minimum national emergency response standards.
These are useful. They are also, every one of them, plumbing.
A hotline is a way to report a frozen car faster. A response guide is a manual for cleaning up after the car has already failed. Geofencing keeps cars out of a zone that first responders have already identified as a problem. None of it touches the moment of failure itself — the instant the car's power dies, its map goes dark, or its planner hits something it has never seen.
That distinction is the whole story, because the failures on the record live entirely outside the crash-rate universe.
Consider the San Francisco Fire Department's actual catalog, as recounted by Fire Chief Dean Crispen and Bay City News. On July 4, a Waymo drove over active fireworks in the street and disabled itself. Other cars reportedly stalled when people touched them with their hands. During December's citywide blackout, Waymos became confused by darkened traffic signals and froze — nearly 1,600 stalls in San Francisco last December, reportedly per the Chronicle.
None of these is a driving error. A car that drives over fireworks and bricks itself has not misjudged a lane change. A car that freezes when the grid goes down has not failed to yield. These are system-reliability failures — power, perception, and fallback behavior under conditions the crash study never measured because the crash study measured miles driven normally.
This is why the 68% figure functions as misdirection even when nobody intends it that way. It is a real answer to a question no one at Fire Station 4 was asking.
Then comes the part that should end the "it's a San Francisco problem" framing for good. In April, in Wuhan, at least 100 Baidu Apollo Go robotaxis reportedly stopped in the middle of the road. Police called it a "system failure." Some passengers were reportedly trapped nearly two hours — the doors opened, but riders were afraid to step out into heavy traffic.
At least 100 cars, one failure, reportedly. That is not a driving-skill deficit that more miles or better training will grind away. That is a fleet-wide reliability event — the kind of correlated, single-point-of-failure collapse that no per-mile crash statistic can even represent, because it isn't about the average mile. It's about the worst minute.
So the correct standard is obvious, and the bill's supporters keep circling it without landing on it.
Cathy Chase of Advocates for Highway and Auto Safety said driverless cars should be compared to "the best drivers, not the worst, or average drivers," and warned that because these failures haven't yet produced a major catastrophe, "we cannot and must not rely on luck." That is exactly right, and it points at the missing requirement. Luck is what you rely on when you have no published test of the failure case.
Here is the gap, stated plainly. The bill demands that operators document stalls, blocks, and collisions. Mullin's earlier bill pushed for more data disclosure of exactly these incidents. But documenting failures after they happen is not the same as proving, before deployment, that the car behaves acceptably when the power dies, the battery drains, the signals go dark, or the central system drops offline.
Nowhere in the described legislation is there a requirement for an auditable, published pass/fail result for those specific scenarios. What does a Waymo do in a total blackout — provably, on a test track, with results anyone can inspect? What is the defined safe state for a dead battery mid-intersection? How long until a system-outage fleet reaches a controlled stop off the roadway? The bill asks for none of this. It asks for a phone number and a manual.
This is where the industry's substitute for proof lives. Waymo's Tiffany Moore, in nearly identical statements to every outlet, cited the company's first-responder program: established in 2018, more than 11,000 first responders trained in California. That is genuine work. It is also, structurally, the answer to the wrong question again — it describes how humans clean up after the car fails, not whether the car fails less.
"We trained 11,000 first responders" is the emergency-scene equivalent of "68% fewer crashes." Both are true. Both measure competence at the edges of the problem while leaving the center — what the machine does at the moment of failure — undocumented.
Notice, too, who is conspicuously silent. Mission Local reached the California PUC and DMV, the agencies that actually hold regulatory power over AVs statewide; neither responded to Lurie's calls for tighter rules or Mullin's bill. Lurie himself, who welcomed Waymo onto car-free Market Street in 2025, has "little authority" here. The bill routes around that vacuum by handing the real technical mandate to NHTSA.
Which is where the entire thing now rests. Everything load-bearing in this legislation is deferred to NHTSA's "minimum standards." The hotline and the geofence are written into the bill; the actual definition of readiness is a rulemaking that hasn't happened yet.
So the test for the coming months is narrow and specific. When NHTSA writes those minimum standards, do they specify a published, auditable behavior for blackouts, dead batteries, and system outages — a defined safe state, a maximum time-to-stop, a result an outside party can verify? Or do they codify the plumbing and call it done?
A named federal bill now exists that could force exactly that. It doesn't — not yet, not as written.
Watch the rulemaking. If it requires proof of the failure case, this becomes a safety bill. If it doesn't, then the Waymo that drifted across the frame behind Lurie was not an interruption of the announcement. It was the announcement.