[ben_w]

Have you never had a speeding ticket from an automatic camera?

Any given AI won’t spot general unlawfulness at this point, but they are getting designed to spot specific laws getting broken.

Perhaps I could have phrased my “in general” better.

[YeGoblynQueenne]

>> Have you never had a speeding ticket from an automatic camera?

I don't drive, so no :)

More to the point, figuring out the speed of a moving object is not an AI task- and I don't mean that in the sense of "if it works it's not AI". I mean, really, it's not something AI was ever interested in, presumably because it's not a particularly complicated calculation, given the right equipment.

Generally, AI is interested in problems that demand, how can I put it, unorthodox solutions. Or just very tricky ones.

So the kind of thing I thought you meant was identifying, say, burglars or muggers, from video feeds etc. That sort of thing is not possible yet, certainly not outside controlled conditions ("in the lab").

[ben_w]

> More to the point, figuring out the speed of a moving object is not an AI task- and I don't mean that in the sense of "if it works it's not AI".

Reading the number plate, on the other hand…

I would be surprised if the AI in any worthwhile self driving car couldn’t detect 90% of categories of unlawful road use.

Detecting assaults may be computationlly unreasonable at this point, but there is work on generating 3D meshes which map to all human bodies in a scene, so it’s not unreasonable to draw a line from one to the other. Identifying that a theft has occurred, however, probably can’t be done yet outside carefully controlled conditions. Yet.

(I have been given one speeding ticket, but in error because I had sold the car before the incident).

[YeGoblynQueenne]

>> Reading the number plate, on the other hand…

Yep, that's a honest-to-god AI task :)

>> Detecting assaults may be computationlly unreasonable at this point, but there is work on generating 3D meshes which map to all human bodies in a scene, so it’s not unreasonable to draw a line from one to the other.

I think the closest analogy is pose estimation, where there's quite a bit of work (in particular, I think there's a lot of interest in learning to identify body postures that can lead to a fall, in order to reduce injuries to older people). I don't remember seeing work on identifying criminal intent in particular, though.

My intuition is that it's more than a matter of computational resources and will require some algorithmic advances. But, you never know.

>> (I have been given one speeding ticket, but in error because I had sold the car before the incident).

Automation, eh? :)

[ben_w]

> Automation, eh? :)

Yup. No idea how it went wrong given I only owned it for only a few weeks specially so I could sell it on behalf of my partner after she accidentally moved to America (I have a complicated reality [1]), and the DVLA sent me acknowledgment, about nine months before I got the letter from the police, that I had sold it.

[1] “And that’s how I found out that Michael Jackson works for the USAF.”