Self-Driving Cars Would Prevent About a Third of Crashes, Not 94 Percent

The 94 percent claim, and what replaced it

You will see it everywhere: human error causes 94 percent of crashes, therefore self-driving cars eliminate 94 percent of crashes. The first half is roughly right. The second half does not follow.

The Insurance Institute for Highway Safety examined 5,000 police-reported crashes and sorted the errors by type. Sensing and perceiving failures, meaning the driver did not see something, accounted for about 23 percent. Driver incapacitation accounted for about 10 percent. Those are the categories a machine reliably designs out.

The largest category, at about 40 percent, was planning and deciding: speeding, aggressive manoeuvres, misjudged gaps. A self-driving car programmed to move through traffic the way people expect will make some of those same choices. The institute's conclusion was that roughly a third of crashes would be prevented unless the vehicles are explicitly built to prioritise safety over speed, which is a design and regulatory decision rather than a technical inevitability.

The safety numbers come from the companies

This is the part that should give everyone pause. Federal regulators require automated vehicle operators to report crashes. They do not require them to report how many miles they drove.

That means the denominator in every safety ratio you have read is supplied by the company being assessed. The regulator says so in its own words, noting that its data have not been adjusted by any measure of exposure and warning against comparing operators using it.

There is independent work emerging, and it is worth watching because it tends to land below the companies' own claims while still showing an improvement. It also comes with caveats that matter: robotaxi fleets largely avoid highways, a large share of their miles are on low-speed streets, and performance varies significantly by city. None of that makes the technology unsafe. It makes the confident numbers unreliable.

There is also very little public data on the technology's failure modes in unusual conditions. Municipalities have documented driverless vehicles stalling in intersections, drifting into emergency scenes and failing to respond to the hand signals police and firefighters use to redirect traffic. Those incidents are recorded locally and inconsistently, which means the national picture of how these vehicles behave around emergency responders does not really exist yet.

Standards written for a driver who is not there

Federal motor vehicle safety standards date from 1966 and assume a person sits behind a wheel. They require rear visibility for that person, airbag warning labels aimed at a driver's seat, and controls positioned for human reach.

A vehicle built without a steering column can be safer while failing rules written for one. That is the strongest technical argument for federal reform, and it is not really contested.

How that reform is happening is contested. In July 2026 regulators allowed one company to build up to 2,500 vehicles a year through 2028 that do not meet portions of eight federal safety standards, including brake systems, rear visibility and interior impact protection. It was the first such exemption for a passenger-carrying automated vehicle, and it came through an exemption process rather than a public rulemaking with notice and comment. Standards were waived before a replacement standard existed.

The coalitions are not what you would guess

Democratic governors in both California and Colorado vetoed bills requiring a human operator in autonomous trucks, against the wishes of their own labour base. Meanwhile Alaska's House passed a human-operator bill unanimously, and similar bills were introduced in Texas, Indiana, Alabama and Iowa, all with Republican majorities.

The reason is that this issue contains two different fights. One is about robotaxis in cities, where the argument runs along familiar lines about regulation and innovation. The other is about driverless freight, where owner-operators and rural legislators see venture-funded fleets running 80,000 pound trucks that never sleep and never take a wage, competing against family businesses with one or two rigs.

Conservatives who distrust concentrated corporate power land in the same place as the Teamsters on that second fight, and federal preemption would override both.

The core disagreement

One argument says delay has a body count. More than 36,000 people died on American roads in 2025. If automated systems are even modestly better than the average human driver, every year spent deliberating costs lives that would otherwise have been saved. Vehicle design has been federal since 1966, and fifty state codes fragment the market without making anyone safer.

The other argument says the evidence is not yet good enough to justify moving fast, precisely because the people supplying the evidence are the people selling the product. It also points at what preemption does to the courtroom: injured people currently sue in state court, where discovery can force a manufacturer to produce sensor logs and internal safety debates, and a single federal standard risks turning that into a compliance defence.

Public opinion is a long way behind both. Asked in February 2026 how comfortable they would be riding in a driverless car, 7 percent of Americans said extremely or very comfortable. Seventy-one percent said not too or not at all, with 43 percent saying not at all. Whatever the engineering shows, the politics of this will be decided by people who do not want to get in one yet.