Leasing the right not to be sued

A friend drove us north through Ontario this summer, and somewhere up there we hit ten kilometres of stopped traffic. Twenty minutes of crawling later the cause turned out to be a short stretch of one lane closed for maintenance, with a crew working in it. We had spent the morning on badly maintained roads, so no complaint about the cones.

Whatever that closure physically took off the road, it was not ten kilometres of it.

I was not driving, so I had nothing to do but watch the rest of it happen. People start late and then start slowly, so a green wave dies three cars in. They hold a gap far larger than they need, or sit close enough that they have to keep dabbing the brake. And one brake light halfway up the line sends a stop back through everyone behind it, long after whatever caused it has gone.

You get more cars past a point per hour in one of two ways. You can pour more road, which costs a fortune up front and then costs again every year forever. Or you can recover the capacity that human reaction time is throwing away. I wanted to know how much of the second one there is.

The number everyone quotes

The Federal Highway Administration would file my jam under work zones. Its national split puts work zones at 10 percent, traffic incidents at 25, and bottlenecks at 40. The category names whatever triggered the queue and then charges it for the entire queue, so a lane closure gets debited ten kilometres of stopped cars when it physically displaced a small fraction of that. The rest is what the drivers did with the disturbance, and no slice of that pie chart holds it.

We crawled past this one later the same trip, and it bills to a different slice with the same error.

The number itself is shakier than the accounting, too. FHWA’s own 2016 review traces the number to a study built mostly on I-95 Corridor Coalition data, notes that the agency has been repeating it for at least fifteen years, and concludes that “no confirmation or additional study has been made since.”

Jams with nothing in the way

Sugiyama and colleagues removed the trigger entirely. They put twenty-two cars on a circular track, asked the drivers to hold a steady speed, and filmed from the middle. Within about three minutes a jam of five cars had formed and was travelling backwards around the circuit as a solitary wave, holding its size and its speed. No merge, no lane drop, no cones, nothing in the way. They conclude that a bottleneck only ever triggers a jam, and that the jam is what a dense line of human-controlled particles does on its own.

So my ten kilometres had a cause I could see and an amplifier I could not. Once a queue forms at a bottleneck, it discharges cars more slowly than it managed minutes earlier. Chung, Rudjanakanoknad and Cassidy measured that drop at 5 to 18 percent across three of them. The road gets worse at the exact moment you need it most, because of how people drive out of a queue.

The gaps have a floor for the same reason. Marc Green’s review of driver braking studies puts perception-brake time at 0.70 to 0.75 seconds when you know the signal is coming, about 1.25 seconds for something ordinary like the car ahead lighting its brakes, and roughly 1.5 seconds for a genuine surprise. At highway speed that is tens of metres of empty road per car. Training does not shorten it; it is how long the wetware takes.

What removing the driver buys

Tientrakool, Ho and Maxemchuk modelled what happens when you shorten that gap. Vehicles using sensors alone raise highway capacity by about 43 percent. Vehicles that also talk to each other raise it by about 273 percent, because they can pull away together instead of one after another. Shladover, Su and Lu found the same shape from real driver gap data: plain adaptive cruise control barely moves lane capacity, and the cooperative version only pays once market penetration is high. Both are models rather than measurements, and I hold them loosely.

The cheap part is already real. Stern and colleagues rebuilt Sugiyama’s ring with one computer-controlled car among about twenty and found that controlling that single vehicle damped the wave for everyone behind it, cutting fuel use substantially. One car in twenty is a deployment target, and it kills precisely the thing I sat in.

To make a lane carry several times what it carries now, you need near-universal adoption and cars that communicate, which means essentially nobody driving manually. That is a ban, and a ban is a legal instrument.

Paying more to fewer people

Ban manual driving and every crash stops being a question about a person and becomes a question about a product. Marchant and Lindor set out the consequence in 2012: liability moves from the driver’s negligence to the manufacturer’s design. That is slower and dearer to prove, and it puts an ordinary plaintiff against a standing legal department. The obvious answer is to stop requiring proof of fault at all, pay out of a central fund financed by a levy on autonomous miles, and take away the right to sue. The United States already runs that machinery for vaccine injuries.

Swiss Re compared Waymo’s liability claims across 25.3 million driverless miles against human baselines drawn from more than 500,000 claims and found 88 percent fewer property damage claims and 92 percent fewer bodily injury claims. The gap holds even against 2018 to 2021 cars with modern collision avoidance. Hold society’s total spending on motor injury compensation flat while claims fall by that much, and every remaining victim can be paid several times what they would collect today, sooner and without hiring anyone.

Why nobody takes the deal

Uber’s test fleet in Tempe was involved in 37 crashes and incidents in autonomous mode between September 2016 and March 2018, according to the company’s own records, given to the NTSB. In 33 of them another vehicle hit the Uber car, 25 of those from behind. Almost none of this reached anyone. Then one of those cars killed Elaine Herzberg, and that crash is what the technology means to most people who have heard of it. Penmetsa and colleagues tracked 1.7 million tweets around that March and a Tesla fatality the same month, and found negative sentiment toward self-driving technology rising from 14 percent to 46 percent in fifteen days.

That asymmetry is not a media failure to be corrected. Koehler and Gershoff gave people cases where the thing meant to protect you is the thing that harms you, like an airbag that kills rather than a crash that kills. Across five studies, people punished the betrayal harder than identical harm carrying no promise of protection. They also chose safety devices with higher overall risk to avoid a 0.01 percent chance of being betrayed by one. Dietvorst, Simmons and Massey found the matching asymmetry in confidence: people who watch a forecaster err drop an algorithm faster than they drop a human who made the same mistake, even after watching the algorithm win overall.

So the bar sits well above parity. Liu, Yang and Xu asked 499 people where the line was, and got four to five times safer than a human driver before the risk read as acceptable. Nees found why it cannot be met by argument: most drivers rate themselves above average, and they want a car safer than they believe themselves to be. A benchmark set against the average driver is therefore one almost nobody applies to themselves. The target moves as you approach it.

Money cannot buy its way past this, and a flat fund makes it worse. It pays by a fixed schedule, and a schedule underpays the people who lose the most, who are also the people best equipped to campaign against it. Worse, a permanent shield priced by a fund turns a death rate into a line item. A company that has bought immunity outright can rationally decline to fix a known defect whenever the fix costs more than the payouts, and everyone can see that from outside.

Renting the shield

So the shield should not be sold. It should be leased, and the rent should be paid in safety.

A manufacturer holds immunity from ordinary product liability only while it keeps cutting its own crash rate against a published baseline every year. Miss three years running and the immunity lapses, the fleet drops back into the tort system, and the accumulated exposure is enough to end the company. The moral complaint dissolves, because no firm under this rule can describe its current death rate as acceptable. The most it can say is that next year’s will be lower, which is a sentence the public will take.

The mechanism is not speculative. The 1970 Clean Air Act demanded 90 percent cuts in tailpipe hydrocarbons and carbon monoxide by 1975. Nobody knew how to build that car, and the penalty was $10,000 for every non-complying one sold, when the average car cost about $5,000. Lee Iacocca told Congress the industry would simply stop building cars. Congress passed it, granted a one-year extension, and the catalytic converter shipped in 1975.

A compounding target cannot run forever. Cutting crashes by a fifth is straightforward while the failures are sensor bugs, and brutal once they are freak events. A regulator that keeps demanding the fifth eventually gets a fleet that brakes for shadows and crawls through intersections, handing back the capacity the exercise was meant to buy. Once a fleet holds its record at some agreed multiple of the human baseline for a few consecutive years, the ratchet stops and the immunity becomes permanent.

Has anyone proposed this

Parts of it.

Conditional immunity is not new. Geistfeld’s roadmap for autonomous vehicles lets a manufacturer discharge its design obligation outright if premarket testing shows the fleet performs at least twice as safely as conventional vehicles. That is a safe harbour bought outright, once. Funkhouser had already argued for the vaccine-court style fund. Technology-forcing regulation backed by an existential penalty is what the Clean Air Act did, and what fuel economy standards do in gentler form. I could not find the combination: a shield whose renewal turns on a rate of improvement rather than on clearing a fixed line once. That is a gap in what I searched. Whether it is a gap in the literature, I do not know.

Two things I believed at the start did not survive the reading. The first is that the entry bar should be high. Kalra and Groves ran hundreds of scenarios and found that deploying at 10 percent better than the average human driver saves far more lives over thirty years than waiting for 75 or 90 percent better, because the waiting is itself lethal. Under a lease the entry threshold should be low and the ratchet steep. The second is the claim that the public needs a hundredfold improvement before it will surrender the right to sue. I went looking for the survey behind that number and could not find one.

What it does not fix

Suppose all of it works. Duranton and Turner used the 1947 interstate plan as an instrument across 228 metropolitan areas and found that vehicle kilometres travelled rise roughly one for one with lane kilometres, concluding that building roads does not relieve congestion. Capacity recovered from shorter following distances is still capacity, and nothing exempts it from that finding. The extra trips are worth something, since somebody wanted each of them, but the queue at five o’clock stays about as long as it is now.

A car with no driver also has nobody in it to help someone fold a wheelchair, find the door, or work out what is happening when the vehicle stops somewhere unfamiliar. That was a service nobody ever priced, because it came free with the person at the front.

Where this lands

I wanted a number for how much of that queue was the road and how much was us. Mostly us: the closure started the jam and the drivers behind it supplied the other nine and a half kilometres. A fleet that reacts instantly and talks to itself would take the larger part of that back, and induced demand would spend it again on trips somebody wanted. The jam comes back.

The injury claims do not. Ninety percent fewer of them is worth having on its own terms, and the obstacle is a rule about who may be sued. Rules are cheaper to change than physics.

So rent the immunity out instead of granting it. Price it in percentage points off last year’s crash rate, strip it from anyone who misses three years running, and hand it over for good once a fleet has held the agreed multiple of the human rate long enough to be believed. The treadmill needs that ending, or it starts spending the capacity it was meant to buy.

The crew in that lane closure will still be there. The ten kilometres behind them do not have to be.