Are Self-Driving Cars Safer?

Ask whether self-driving cars are safer than human-driven ones and you will get a confident answer. Usually two, pointing in opposite directions.

Both are answering the wrong question.

The interesting divide is not human versus machine. It is between technology that helps a driver and technology that replaces one. Those are not the same project, and folding them together under “self-driving” hides the thing that actually matters.

Start with the good news, because there is a great deal of it and it rarely makes the headlines.

The advanced driver aids quietly fitted to ordinary cars — automatic emergency braking, lane-keeping, blind-spot monitoring, stability control, adaptive cruise — are among the most effective safety improvements the industry has ever shipped. They work not by being cleverer than the driver, but by being awake at the precise half-second the driver isn’t.

Human attention is a flickering thing. It wanders, tires, checks the mirror at the wrong moment and assumes the car ahead will keep moving because it has kept moving for the last hour. The machine’s single advantage is that it never gets bored. It intervenes appropriately: a nudge on the wheel, a stab on the brakes, a warning a fraction of a second before the driver would have noticed the problem themselves.

That is a colossal safety win. Better still, it is available today, in thoroughly unglamorous hatchbacks, without anyone having to solve full autonomy first.

But — and this is the whole argument — it works only while the driver is still driving.

These aids are at their best as a second pair of eyes on a fully engaged human. The moment the technology does enough to convince the driver they are no longer needed, the arrangement quietly inverts. Now the human is the backup. And humans are catastrophically bad at being backups.

We are simply not built to monitor a system that is right ninety-nine times out of a hundred, while staying poised to seize control on the hundredth. Watch a road you are not actively driving on and you will bring roughly the vigilance of someone waiting for a kettle. By the time the alert sounds, the context has gone, your hands are somewhere unhelpful, and the two seconds you need to understand the situation are precisely the two seconds you do not have.

This is the trap of the middle ground: cars competent enough to lull, but not competent enough to trust. “Nearly self-driving” can be more dangerous than either a plain car or a genuinely autonomous one, because it removes the driver’s attention long before it can safely remove the driver.

So my view — and it is a view, held at the appropriate level of confidence — is that the safety case for this technology is real and large, but it points somewhere specific. The gains come from aids that intervene to support a driver who stays firmly in control and firmly responsible. They do not come from gently persuading that driver to stop paying attention.

The distinction is really about abdication.

A good aid catches your mistake and hands the situation — and the responsibility for it — straight back to you. A bad one invites you to hand the responsibility over and simply hope for the best. The first makes a good driver better. The second manufactures a worse one and calls it progress.

Framed that way, a lot of the noisy debate resolves itself. Automatic braking that saves a distracted parent from a child running out is unambiguously good. A system that lets a commuter watch films on the motorway until it panics and flings the wheel back at them is not the next step up from that. It is a different, more dangerous idea wearing the same badge.

None of this is an argument against ambition. Full autonomy, done properly, may eventually be safer than all of us, and I would welcome it. In the narrow, well-mapped, fair-weather cases the best systems are already genuinely impressive. But in the messy long tail — the roadworks, the ambiguous hand signal, the sheep that has decided the far verge looks greener — they are not there yet, and the driver remains the only thing standing between that long tail and the ditch.

Until that changes, the safest car on the road is not the one that drives itself. It is the one that lets an attentive human drive, and quietly rescues them from the half-second they got wrong.

So keep the aids. Keep improving them, aggressively. But keep the driver’s hands on the wheel and their name on the outcome.

A co-pilot makes the pilot better. A chauffeur just teaches the passenger to stop watching the road.

• #thinking
• #safety
• #autonomous-vehicles

@online{byatt2026-areselfdrivingcarssafer,
  author  = {Stephen Byatt},
  title   = {Are Self-Driving Cars Safer?},
  year    = {2026},
  date    = {2026-07-26},
  url     = {https://learningoutloud.stephenbyatt.com/blog/are-self-driving-cars-safer/},
  urldate = {2026-07-26},
  note    = {Learning Out Loud}
}