As claims circulate in the autonomous driving industry that self-driving vehicles are "100 times safer than humans," discussion continues over what those figures rest on.

The first issue is what is being compared. Autonomous driving data is mostly collected in limited areas of specific cities under controlled weather and time-of-day conditions. Human crash statistics cover highways and mountain roads, night driving and bad weather alike.

The second is the definition of a crash. Whether the comparison uses contact incidents, injury crashes or fatal crashes changes the multiple dramatically. Including minor contact worsens the autonomous figures; restricting it to fatalities leaves too small a sample to be statistically meaningful.

The third is attribution of fault. How to count crashes involving an autonomous vehicle where the other driver was at fault is unresolved, and the choice yields different conclusions from identical data.

The general direction is nonetheless widely accepted. Drink-driving, fatigue and distraction — leading causes of human crashes — do not arise in principle for an automated system. The question is under what conditions, and to what degree, that advantage materialises.

Regulators in several countries are converging on defining the operational design domain precisely: documenting which roads, which speeds and which weather conditions a given autonomous function is permitted to operate in.

For consumers the practical test is simple: check the conditions under which the function was approved to operate, and what it requires of the driver when those conditions no longer hold.