Just add two buttons to the app "minor cleanliness issue" and "major cleanliness issue". Then someone remotely checks the footage. If it is confirmed then the previous passenger gets charged a cleaning fee. Majors cost more or could result in ban. Cleaning can be done by carwashes.
Eventually you develop some clever covers and devices/stations that enable rapid cleaning.
With all the computing power and vision AI available in such a vehicle, it seems like it would be possible to perform an automated interior scan of the cabin after each ride. Detecting cleanliness issues(and forgotten baggage).
Which of course creates a self-perpetuating problem, because if the behaviour is not punished, it will never be corrected and it will become "ok" to make a mess in an auto taxi(not vomit, but just leaving your rubbish behind).
Fortunately, I can get that done for me at a cost of much, much less than 12 hours of my own time. I can farm it out overseas to somewhere with a large supply of very cheap labour. I wouldn't be surprised if sitting in front of a bank of a half-dozen taxi-interior replays running backwards at triple-speed (with the empty periods already taken out by some relatively simple processing) from a point at which the uncleanliness is visible was seen as a good job, compared to some other options available.
Wouldn't be surprised if I could have that done for me at a cost of a few dollars a time; if I then charge the culprit a hundred bucks on their already registered credit card or taxi account for cleaning (fifty sounds too low), it becomes very economical for me.
I mean, we have to deal with shitty customers all the time. 99% of the time it's just not worth pursing someone who broke something or just left a mess, even if theoretically I could hunt down proof that they did it. It almost always escalates to some multi-month situation that is not worth the effort.
If someone becomes a multiple offender, you ban them.
Simple.
You know exactly what time your customers entered and left the vehicle. All you have to do is pull out a set of freeze frames from before and after each customer and compare the two.
As others have mentioned, the app can have a button to report a cleanliness issue. Working back from the first report means you don't have to review 12 hours of footage. As a side note (since this is HN), you don't have to review 12 hours linearly, you can use a binary search, which means you can probably find the culprit in minutes at most.
The cars are going to have cameras constantly running in them.