The 115 We Dont Count

Every day in the United States, roughly 115 people die in car crashes caused by human drivers. That number doesn’t make the news. It barely makes the statistics. It sits inside a category called “unintentional injuries” on the government’s list of what kills Americans, so thoroughly normalised that it doesn’t even get its own line item.

On a Saturday night in December 2025, a power substation fire in San Francisco knocked out traffic signals across a third of the city. Waymo’s driverless vehicles did exactly what they were coded to do: they treated dead signals as four-way stops, sometimes pausing to request confirmation from a remote operations team. Videos on social media showed the cars stopped with hazard lights blinking. A city supervisor called for a hearing. Safety experts declared it a warning. Local news covered it extensively.

No one was injured. No cars crashed.

Across more than 220 million miles of driving, Waymo’s vehicles have been involved in 94% fewer crashes causing serious injury or death than human drivers on the same roads. Pedestrian injuries are down 93%. Cyclist crashes, 84%. Intersection crashes, among the deadliest category in trauma medicine, down 96%. If these numbers hold as deployment scales, the population-level health effect could rival seat belts or the decline in smoking.

And yet the dominant public conversation about autonomous vehicles is not “when can we have more of these?” It’s “should we trust them at all?”

That’s not a technology problem. That’s a psychology problem. And it tells us something important about how humans evaluate AI across every domain, not just driving.

The Inflated Mirror

Psychologists call it illusory superiority: the hardwired tendency to believe you are above average. Nearly every driver thinks they’re better than the median. It shows up across countries, across decades, and across every demographic you can measure. It’s not arrogance in the traditional sense. It’s a cognitive default. We experience the world from inside our own decision-making, and from in there, it looks pretty competent.

What this means in practice is that when we evaluate a machine’s performance, we don’t compare it to the actual distribution of human ability. We compare it to the version of ourselves we carry around in our heads, the one that always signals correctly, never gets distracted, and would definitely have noticed the pedestrian. The evidence has to be better than the driver we imagine ourselves to be, not the driver we actually are.

This is why strong safety data hasn’t led to more public trust in autonomous vehicles. The data is being measured against a fantasy.

The Bandage on the Floor

Dr. Thomas Lee, a Harvard professor and internist, spent decades studying what patients actually notice in hospitals. For years, the assumption in medicine was that clinical safety was the whole job: keep infection rates low, get good outcomes, and trust would follow. That assumption was wrong.

A patient comes in for a routine knee replacement. The surgery goes well. Correct implant, clean wound, no complications, discharged on schedule. Two weeks later, she misses her follow-up appointment. A month later, she transfers her care.

What happened? She overheard two nurses arguing in the hallway outside her room the morning after her procedure. She didn’t hear what they were arguing about, but she felt that the people taking care of her weren’t working together. The clinical quality was there. The experience wasn’t.

When Lee’s team ran natural-language processing across patient comments, the trust-killers were rarely clinical errors. A patient wrote about a used bandage on the floor, or a doctor who appeared to have a bloodstain on his scrubs. They knew these things wouldn’t hurt them. As Lee put it: “They were reminders that there is no such thing as perfect safety, but who wants that reminder when they are sick?”

A patient cannot directly observe surgical quality. What they can observe is the floor, the scrubs, and the room. They read those signs and draw conclusions. Healthcare had to learn to treat this kind of semiotic failure, where the signals a system sends diverge from what it actually delivers, as a separate design problem.

The December San Francisco blackout, captured on cellphone cameras, was, in Lee’s framework, the bandage on the floor of a self-driving car. The machine was safe. The signals it sent said otherwise.

No One to Forgive

Here’s where it gets interesting, and where it applies far beyond autonomous vehicles.

Empathy is a cognitive tool we use to forgive human error. When a person makes a serious mistake, we have somewhere to place the blame and, eventually, someone to absolve. A surgeon who loses a patient has a face, a voice, a story. We can imagine the difficulty of the decision. We can project ourselves into the moment. System 1 does this automatically: we feel the error as human because we are human.

A machine’s error is different. There’s no agent to forgive, no emotional arc to follow. It’s the product of code, of corporate decisions, of systems so complex that no single person can be held responsible. Behavioural scientists call it algorithm aversion: people tend to lose confidence in a machine more quickly than in a human when the two make identical mistakes. We prefer a worse human over a better algorithm if the latter has made even one visible error.

This is the double standard at its most lethal. We will tolerate 115 deaths a day from human drivers because we can empathise with the humans who caused them. We will demand near-perfection from autonomous vehicles because we cannot forgive them the way we forgive each other.

And it’s not just driving.

The Same Pattern, Everywhere

Most Americans say they would be uncomfortable if their doctor relied on AI to help diagnose them. But patients who actually receive AI-assisted screening report satisfaction levels above 90%. People who object most loudly to self-driving cars haven’t been in one. Confidence in the technology runs far higher among riders than non-riders.

The pattern is consistent: the people with direct experience trust the technology. The people without it don’t. And the people without it are making the decisions.

This is the real cost of the perfection standard. It’s not that AI is imperfect. It’s that we’re holding it to a standard we have never applied to ourselves, and using that standard to delay the adoption of things that would save lives.

In healthcare, AI diagnostic tools already outperform the average clinician in specific tasks: radiology screening, dermatological lesion classification, diabetic retinopathy detection. The evidence is not ambiguous. But the conversation is never “AI is better than the average doctor.” It’s “AI missed this one case,” presented as though the average doctor doesn’t miss cases every single day.

The denominator is always missing. We count AI errors. We don’t count the human errors they replace.

The Honest Reckoning

The Noema article’s author, Jonathan Slotkin, is a neurosurgeon. He makes a point that cuts through the entire debate: “The problem isn’t a performance problem; it’s an interpretation problem.”

That’s the line that should keep anyone working in AI, in marketing, in any field where AI is being deployed, awake at night. Because the technology can be objectively, measurably, overwhelmingly better than the alternative, and it won’t matter if the interpretation is wrong.

The Waymo that stopped at a dumpster in San Francisco did exactly what it was supposed to do. It identified an obstacle, flagged it, and waited for resolution. The machine performed correctly. But the human in the back seat, a neurosurgeon who had just written a New York Times op-ed arguing for the technology, grabbed his phone to record what wasn’t happening. “The machine had done what it was supposed to do,” he wrote, “but it was still hard to shake the feeling that it was broken.”

That feeling is the entire problem. Not the technology. The feeling.

Every major safety technology follows the same arc: initial resistance, peak anxiety, gradual exposure, eventual normalisation. Seatbelts were opposed as government overreach. Anti-lock brakes were distrusted because the pedal feedback felt wrong. Airbags, in their first generation, actually killed people, and we still adopted them because the alternative was worse.

AI is on the same arc. The question is how many people die while we’re waiting for the cultural default to flip.

What This Means for Anyone Working With AI

None of this means AI is without problems. Cruise, another autonomous vehicle company, had its licence suspended after one of its cars dragged a pedestrian 20 feet and the company’s report to regulators omitted that detail. That’s not a perception problem. That’s a trust problem earned through dishonesty.

AI systems hallucinate. They make confident assertions about things that aren’t true. They encode biases from their training data. These are real issues that require real solutions. But they are not reasons to reject the technology wholesale, any more than the first-generation airbag deaths were a reason to remove airbags from cars.

The honest position is this: AI is imperfect, and it is better than the average human at an increasing number of tasks. Both of those things are true simultaneously. The question is not “is AI perfect?” because nothing is. The question is “is AI better than what we have now?” And in case after case, the answer is yes.

We demand perfection from machines because we cannot empathise with them. We tolerate carnage from humans because we can. That’s not a rational safety calculation. It’s a cognitive bias, and it’s killing people.

The 115 we don’t count every day are the price of our inflated mirror.