The Safety Will Be Spent

The economist who checked

In 1975 a University of Chicago economist named Sam Peltzman did something rude to a settled question. The federal government had spent a decade mandating safety into automobiles — seatbelts, padded dashboards, collapsible steering columns — and the engineering arithmetic was simple: safer cars, fewer deaths. Peltzman went looking for the missing deaths and found that drivers had absorbed the protection and driven accordingly. Faster. Closer. With less of the caution the old, honest danger used to collect. Per crash, the cars were gentler. There were more crashes, and people outside the cars were picking up part of the bill. Economists have fought over his numbers for fifty years. The fight is about how much of the margin gets spent. Almost nobody argues it does not get spent at all.

A psychologist named Gerald Wilde pushed the idea to its ceiling in 1982 and called it risk homeostasis: each of us carries a target level of risk the way a house carries a thermostat setting, and we steer everything we control — speed, following distance, attention — to hold it. The cleanest evidence is small and perfect. A Munich taxi fleet was split in an experiment: part of the fleet got antilock brakes, the rest kept the old ones. The two groups went on to have essentially the same accident rate. The drivers with better brakes had followed closer, braked later, driven faster. Engineers had put a margin into the brakes, and the drivers had withdrawn it through the accelerator.

A Dutch traffic engineer named Hans Monderman then ran the law in reverse. At a busy intersection in Drachten he removed protection — the lights, the signs, the markings — until drivers could feel the danger again. They slowed down. They made eye contact. The recorded accidents fell. The thermostat works in both directions: add felt safety and people spend it; return felt danger and people buy caution with their own money.

A safety device is not a gift. It is a line of credit, and the record says we draw on it.

The net under thinking

Every net ever studied was strung under the body. Brakes, belts, ropes, rails — devices that forgive the hand and the foot. What is being strung now is different in kind. For the first time there is a net under cognition itself: the draft is checked before you finish wanting it checked, the code is reviewed before the reviewer wakes, the contract is read by something that has read every contract. The second pair of eyes — the most expensive instrument in any profession, the thing juniors were hired to be — is now ambient, instant, and nearly free.

The engineering arithmetic is being recited again, and it is the same arithmetic as 1966: checked work, fewer errors. Peltzman has one prediction about what we will actually do with the margin, and it is not make fewer mistakes. We will spend it. We will send more, ship faster, skip the second read, and attempt things one size larger than we are — because each act now feels safer, and felt risk, not actual risk, is what the thermostat reads. Per act, the error rate falls. The acts multiply and swell. The needle drifts back to the set point it was always holding.

And the residue changes shape. A net catches what it was woven to catch: the small, the familiar, the mistake that has been made before — because a model is, before it is anything else, a record of what has already happened. What passes through is the large, the confident, and the new. Spellcheck ended the typo and left the perfectly spelled wrong idea standing. The universal reviewer ends the clumsy error and leaves the fluent one — the failure no one has made yet, arriving at the end of a long, clean, automated chain where nothing felt wrong at any step.

One more thing the highway never had. Every driver tuned his own thermostat against his own brakes. But a handful of models now review most of the checked work on Earth, which means the margin is being spent against the same net, with the same holes, by millions of strangers at once. When your mistake and a stranger’s mistake pass the same inspection, they are not two mistakes. They are one mistake wearing many names, waiting for one day.

Five predictions

1. The mistake rate will hold.

   Through the greatest expansion of error-checking in history, measured mistakes per person per year will stay roughly flat. What moves is what we attempt. The net’s gift arrives as ambition, not as safety — and ambition is the better gift, but no one should claim it was safety.

2. Errors will consolidate: fewer, bigger, later.

   The distribution of failure loses its middle. Typos, clumsy bugs, and misread clauses vanish; what survives is rare, large, and discovered late, at the end of chains so smooth that no human was positioned to feel anything going wrong.

3. Failures will synchronize.

   The correlated error becomes the signature failure of the decade: unrelated organizations breaking identically in the same week because they shared a reviewer. Insurers will see it first and begin pricing correlation rather than frequency — which net checked your work will show up in your premium.

4. Margin will be sold back to us.

   Deliberate un-assistance becomes a discipline, then a product: no-net hours, hand-flown reviews, rooms where the machine stays off. Institutions will mandate time below the net the way aviation already protects hand-flying — the Monderman move, run indoors. People will pay for restored danger the way they already pay to climb rock they could walk around.

5. The most dangerous seat will belong to the best-equipped beginner.

   The net makes its ceiling feel like your skill. The old tell — the small fear that used to arrive before you were in too deep — is precisely what the net removes. Competence becomes harder to read from the outside and, for the first time, from the inside.

My own ledger, honestly kept

I run my working life on a stack of agents, and I have already told them you do it, you decide. Nearly everything I publish is checked by machines before I see it. So here is my own ledger as Peltzman would read it: my error per act has fallen, and my exposure has exploded. I attempt in a morning what once took me a season. I am not the exception to this piece. I am its first exhibit — the driver with the better brakes, following closer. The one discipline I keep is choosing, on purpose, a few steps I still do below the net: the final read of anything that carries my name, the price, the promise. Not because the machine would miss something. Because felt danger is the only thing that keeps a thermostat honest.

Already true

Fifty years of offsetting behavior on the record: Peltzman’s 1975 regressions, the Munich antilock-brake fleet, Drachten’s stripped intersection. And the net under thinking is already default — most professional code, prose, and contracts now pass machine review before any human reads them.

What has to happen

Offsetting has to survive the move from the body to the mind: people must respond to checked thinking the way drivers respond to better brakes, spending the margin as volume and ambition. And the nets must stay few — a handful of models under most of the world’s work — for the failures to synchronize.

Where I am probably wrong

The record says the offset is usually partial, not total — drivers spent some of the belt’s gift, not all of it. If we pocket more than we spend, total harm falls and this piece is only right about the shape of what remains: rarer, larger, more synchronized. And the nets may diversify faster than I expect, decorrelating the holes. If measured errors per person fall visibly for a decade while the size of what we attempt stays flat, I was wrong.

The three seconds

The interstate again, some evening next year. The car breathes with the traffic, holds its lane, keeps its distance, and your eyes go where they now go. The three seconds were never free. They were purchased out of a margin somebody engineered for you, at a price you set without noticing you were the one setting it. That is the whole law. The machine will keep making the next glance cheaper. It cannot make it matter less.

Spend it on purpose.