I don't get what this means. If you use basic statistical methods and realize that murders are more likely to occur in this location between these hours and adjust your police force accordingly, is this bad? Shouldn't police use basic statistical methods in determining effective policies?
It depends - These models are often non-transparent in terms of how they calculate, so what if the AI/model decides/calculate that a higher percentage of minorities in an area is a high predictor of crime, thus indirectly creates a policy that, for instance, areas with high levels of ethnic minorities require higher levels of policing or harsher restrictions (i.e. curfews, more CCTV e.t.c.).
If it's just a basic 'well murders happen between these hours on this street so we will put a police officer there' then of course that should happen, but that's not really policing based on AI/ML, it's just based on basic statistics.
For instance, decision trees can tell us if the population density between this range and the percentage of single family residence is between this range, the probability of a package being stolen is X.
Or regression could tell us that the square of the population density is positively correlated to this crime .
It's pretty simple to make the models transparent. Just select the inputs you want the model to have access to. You shouldn't be feeding your model the percentage of minorities in an area.
What's not transparent is how humans make decisions. You can't tell a human to consider only these characteristics but not these others ones.
> If it's just a basic 'well murders happen between these hours on this street so we will put a police officer there' then of course that should happen, but that's not really policing based on AI/ML, it's just based on basic statistics.
I don't see how you make a meaningful separation between the two. Either you can use predictive models or not. It's hard to carve out "use predictive models but not too predictive". The more likely scenario is just not using any model and distributing police resources uniformly regardless of need.
That doesn't make the model transparent, it just makes the input and output transparent, but it's still a black box.
It's nice to say that everyone will ensure that the model isn't fed with information which will allow it to make decisions or be biased based on sex / gender / race / income but in reality accidental racial bias has already shown up multiple times in applications where it should be easier to detect than an invisible 'crime prediction' algorithm that the general public won't have access to but public policy will be implemented on the basis of.
> I don't see how you make a meaningful separation between the two. Either you can use predictive models or not. It's hard to carve out "use predictive models but not too predictive". The more likely scenario is just not using any model and distributing police resources uniformly regardless of need.
I make a meaningful separation because one has clear logic which can be viewed to be unbiased, and the other has a black box calculating it.
> accidental racial bias has already shown up multiple times in applications where it should be easier to detect than an invisible 'crime prediction' algorithm that the general public won't have access to but public policy will be implemented on the basis of.
The primary victims of violence are often of this same race as the "accidental racial bias" you're talking about. You're not doing anyone any favors by misallocating police resource to be more "equitable" without considering actual need. Only 7% of Americans want less policing in their neighborhoods. I imagine people that live in high crime areas would prefer more policing.
https://thehill.com/hilltv/what-americas-thinking/562738-pol...
you’re on the wrong track if murders (and the like) are the crimes you’re targeting. murders, while horrific, are a tiny and very localized danger. the crimes we need to be policing more actively happen around money and power (mass surveillance, corruption, etc.), which are far too common and have broad, highly externalized, and surreptitious effects.
i’d heartily support (heavy-handedly, even) applying any and all such techniques on (so-called) white-collar crimes. in contrast, social norming in communities (like not murdering) should be convivial rather than adversarial, as that’s much more effective (as evidenced by how little murder actually happens).
Because they are naturally uncommon and localized or because they are policed heavily?
If police leave 45% of murders unaddressed, I wonder how many current murders are secondary in nature, as in would not have occurred without an unretributed precipitating crime be-it murder, rape, assault, etc.
It’s an interesting idea to think about murder not being policed or have legal repercussions. Would it be comforting to know if you kill somebody drunk driving there wouldn’t be a big problem? I guess somebody like Dallas’ Dr. Death would still have to deal with pesky civil lawsuits.
we'd be better served by having many more detectives, and far fewer gun-toting beat cops, if we were interested in solving more murder cases though. that's the kind of policy discussion we should be having, not the politically-driven narratives that currently prevail around the topic.
How would that be different from vigilantes and lunch mobs?
What would the ideal ratio of police to detectives be in different contexts?
For instance, I lived in an area where there was a lot of bike thefts. Everyone knew it and police got a lot of bike theft reports (it was in the statistics). I would like to see more police enforcement to deter bike thefts. I think its unethical to ignore the needs of a neighborhood when allocating police resources.
No. For one, police are not statisticians. Here's a military example from WWII:
During World War II, the statistician Abraham Wald took survivorship bias into his calculations when considering how to minimize bomber losses to enemy fire. The Statistical Research Group (SRG) at Columbia University, which Wald was a part of, examined the damage done to aircraft that had returned from missions and recommended adding armor to the areas that showed the least damage. This contradicted the US military's conclusion that the most-hit areas of the plane needed additional armor. Wald noted that the military only considered the aircraft that had survived their missions – ignoring any bombers that had been shot down or otherwise lost, and thus also been rendered unavailable for assessment. The bullet holes in the returning aircraft represented areas where a bomber could take damage and still fly well enough to return safely to base. Therefore, Wald proposed that the Navy reinforce areas where the returning aircraft were unscathed, inferring that planes hit in those areas were the ones most likely to be lost. His work is considered seminal in the then-nascent discipline of operational research.