Removing the driver removes the error behind most collisions

The argument

The advantages claimed for self-driving cars begin with what goes wrong in ordinary driving. Proponents argue that collisions are overwhelmingly the product of the person in the seat rather than of mechanical failure: inattention, misjudged distance and speed, fatigue, intoxication, impatience and the ordinary lapses of someone performing a demanding task for the thousandth time. A system does not become bored, does not drink, does not look at a telephone and does not misjudge how long it has been awake. Proponents add that the machine's advantages are not only negative ones. Sensors can watch in every direction at once, hold a consistent following distance, measure closing speed rather than estimate it, and react without the delay a person needs in order to notice, decide and move. A fleet also learns differently from a population of drivers: where a person learns from their own near misses, a vehicle can be updated from situations encountered anywhere in the fleet, so a hazard handled badly once can be handled better everywhere afterwards. The claim is therefore about the shape of the risk rather than about perfection. Proponents accept that automated vehicles will have collisions, including some a person would have avoided, and argue that the comparison which matters is with the existing toll on the roads rather than with an ideal. On this account the advantage is structural: the commonest causes of harm in traffic are features of human attention, and a technology that takes the human out of the loop removes those causes instead of mitigating them.

Premises

Counter-arguments

Critics reply that the comparison is being made against the wrong baseline. Human drivers accumulate an enormous record of uneventful journeys, and matching it is harder than the list of human failings suggests, because the difficulty of driving lies in rare and ambiguous situations rather than in the routine. They add that removing human error introduces machine error of a different kind: sensors degraded by weather, software that behaves in unexpected ways, and failures correlated across a whole fleet rather than confined to one driver. A mistake made by one person injures the people nearby, whereas a mistake in shared code can be made everywhere at once.

Rejecting the premises

[Rejecting P1] Attributing a collision to the driver does not establish that a system would have avoided it, since many involve ambiguous situations that are hard for any agent, human or automated. [Rejecting P2] Machine attention has its own failure modes, including sensors degraded by weather and software behaving unexpectedly, so the comparison is between kinds of error rather than between error and none. [Rejecting P3] Fleet-wide updating propagates faults as readily as improvements, so the mechanism the premise relies on can place a defect in every vehicle at once. [Rejecting C] At most it follows that automation removes a familiar class of error, not that the overall rate of harm falls.