Should we build the sensors that let a car count the occupants of the vehicle it is about to hit?

Only if the number is allowed to matter — the normative question comes first, and if offsetting lives is impermissible then the technology to count them is unnecessary.

* The normative question comes first — a capability you may not use is not a neutral asset *

The technical questions are real and unsettled. Will the car reliably distinguish people from other objects? Will it be able to tell how many people are inside another vehicle? Will it distinguish a woman with a pram from a child with a toy pram? How a car can be programmed will depend on what is technically possible.

But the paper makes a methodological point that is easy to get backwards: the prior question is what is desirable and necessary at all.

• Should it matter to an autonomous car's behaviour how many people are sitting in another vehicle — or would "offsetting human lives" (Aufrechnen von Menschenleben) constitute an unacceptable instrumentalisation of those sacrificed?
• If the latter, then technology that counts the occupants of the other car is possibly unnecessary.
• If the number of endangered people may play no role from a normative point of view, there is no reason to make the corresponding distinction technically possible at all.

This inverts the usual order of engineering discussion. The instinct is to build the capability and let policy catch up; the argument here is that a capability you are forbidden to use is not a neutral asset — it is a standing temptation and a liability, and the requirement should be derived from the norm rather than the other way round.

Tip: This is a general design principle, not a car-specific one. Data you may not act on is data you should not collect — the same logic behind data minimisation in privacy law.

Go deeper:

• Wikipedia: Machine ethics — the field that asks which decisions should be delegated to a machine at all, not merely how.