Hacker News new | past | comments | ask | show | jobs | submit login
Predictive policing lacks accuracy tests (muckrock.com)
55 points by danso 7 days ago | hide | past | web | favorite | 29 comments

not only does it lack accuracy tests, it has been studied in the past, for example by RAND¹ and there's no evidence that it works at all.

more importantly, it should be noted that these are essentially unproven systems that are being beta tested on largely minority groups. That something like this is happening in a country with a rule of law and at least in theory constitutional safeguards is reprehensible.


The study you link is of a very different kind of system: creating a list of people considered likely to be victims of homicide or violent crime, and trying to intercede.

PredPol, the system in the article, does not do that. PredPol is primarily about mapping crime in space, and trying to predict which regions of space are most likely to have future crime. It doesn't focus on individuals in any way.

The people behind PredPol did run a randomized controlled field trial with LAPD and claim to have found reductions in crime by directing patrols to predicted hotspots [1]. Hotspots policing studies generally have found reductions in crime when directing patrols to hotspots, or conducting targeted interventions (like improving street lighting, talking to local businesses, etc.). But I should note that running a good, well-controlled experiment of something like policing is very hard, so the quality of the evidence is not exactly astounding.

However, there are indeed constitutional implications, as you note. For example, if a model says an area is a "high crime area", police may be able to use that as part of the justification for stop-and-frisk, reducing the need for any individual suspicion or reason to search you. And, of course, there are fairness problems: high-crime areas tend to be poor and minority areas, and hotspots policing may harm trust in police (if they harass the residents for petty offenses) and either lead to overpolicing or underreporting of offenses.

I have a fairly thorough review of the literature on my website [2].

[1] https://www.tandfonline.com/doi/abs/10.1080/01621459.2015.10...

[2] https://www.refsmmat.com/notebooks/predictive-policing.html

How does the current system work ? AFAIK, the police chief looks at various data sources like heat maps and police reports. He will have an incentive organize patrols to protect the politically connected and their property. Then his job is safe.

It's pretty well known that increased policing doesn't decrease crime. These are the same tactics in new clothes used to target PoC. Even if the ML is _working properly_ the biased nature of the existing justice system is going to cause it to target PoC disproportionately.

target PoC disproportionately.

But what would be the purpose of it? Are you saying LAPD wants to put all black people in cages and then enjoy life in a city made exclusively of white (and chinese and indians)?

Or they simply enjoy being dickheads and stop-and-frisking people?

I don't know what the purpose is. I can't speak to the intentions of anybody involved, but the data are pretty clear. PoC are disproportionately stopped-and-frisked, they're disproportionately brought to court for minor offenses, disproportionately killed by police, and disproportionately sentenced, relative to white defendants.

The why is an interesting thing to pursue, but I haven't seen a satisfactory explanation of why. The what and how are more clear, though.

>But what would be the purpose of it? Are you saying LAPD wants to put all black people in cages and then enjoy life in a city made exclusively of white (and chinese and indians)?

It's mostly not about the fact that they're minorities. Ppeople have the misfortune of living in places with high enough population density to support this kind of policing just happen to be mostly minorities. The people pushing for Orwellian surveillance and policing of the inner city would be happy to expand it to the land of trees and double wide trailers were that practical to do so within current budgets.

I'm sure there's some racists here and there within government but not enough to be of consequence or to push these kinds of things on their own (IMO).

>Or they simply enjoy being dickheads and stop-and-frisking people?

They see themselves as "preventing crime" but I think it's a glass half-empty/half-full type distinction, not that it matters. It's a violation of people's rights no matter how you cut it.

What if they are run by the local communities?

Or would they prove useful in homogeneous populations with rises in certain kinds of crime?

PoC stands for People of Colour. Took me a while to work that out. I went off on a tangent trying to work out the tactics clothing manufacturers use to increase a metric called PoC.

PoC is Proof of Concept in every context I've ever seen it before.

Acronyms are often overloaded across industries. If you find yourself at a cocktail party with a bunch of biologists you will confuse them less if you say you work on machine learning than AI. A sports MVP is basically the inverse of an engineering MVP. It is no surprise that PoC in technical fields is not the same as PoC in social sciences.

> It's pretty well known that increased policing doesn't decrease crime.

I suspect that statement has a LOT of context behind it--citation please?

I could believe that increased policing simply pushes the crime around rather than gets rid of it.

> These are the same tactics in new clothes used to target PoC.

Agreed, unfortunately.

There's a lot of existing research here. Generally, the studies show that directing police patrols to high-crime areas does reduce reported crimes. For example, a recent systematic review of studies found that increasing policing to high-crime areas "generates small but noteworthy crime reductions, and these crime control benefits diffuse into areas immediately surrounding targeted crime hot spots." [1]

The question of whether policing simply pushes crime to another area is a popular one, and there's been a lot of work to answer it. It appears the answer is usually "no." Crime depends on several factors existing in one place: opportunities for crime, a lack of guardians, and motivated offenders. If you fix one area, the crime will spread to adjacent areas only if those three factors are present there.

The fairness and equity problems are, of course, severe. You may reduce crime with hotspots policing, but as the systematic review notes, "only a small number of studies examine the impacts of hot spots policing on police-community relations," or what disproportionate effects they have on any specific group.

[1] https://doi.org/10.1080/07418825.2012.673632

>I suspect that statement has a LOT of context behind it--citation please?

It's completely false of course. Not only does policing decrease crime, the US is massively underpoliced under any cost-benefit analysis.



'We have demonstrated that predictive policing of drug crimes results in increasingly disproportionate policing of historically over‐policed communities. Over‐policing imposes real costs on these communities. Increased police scrutiny and surveillance have been linked to worsening mental and physical health;and, in the extreme, additional police contact will create additional opportunities for police violence in over‐policed areas.'


Increased police scrutiny and surveillance have been linked to worsening mental and physical health

I find it surprising. I would be happy to see more police in where I live. In my layman's mind police, pretty much by definition, prevents and deters crime. They may not be perfect in that, but in most cases they're fine.

> In my layman's mind police, pretty much by definition, prevents and deters crime.

That would be the case if police were adequately trained and equipped. However US police forces on average (!!) are trained only for not even 9 months (per https://www.snopes.com/fact-check/police-training-killings-u...), which means that there are likely cases in which only weeks of training were given. German police, in contrast, train for two and a half / three years (per https://de.wikipedia.org/wiki/Polizeiausbildung_in_Deutschla...).

That, combined with a lack of hiring standards (e.g. https://www.portlandmercury.com/blogtown/2019/06/19/26671582...) leads to the perception of police being nothing more than paid thugs with a uniform. The massive availability of guns as well as former military equipment in police hands doesn't help either, quite to the contrary. When I have an easy sledgehammer (=a tank and a gun), I'll shoot first and ask questions later.

When it comes to "overpolicing", you also have to note that it's way easier to arrest someone dealing/smoking cannabis on the street (mostly poor PoC) compared to arresting a rich white banker who is snorting cocaine on his desk.

I'm going to speculate that you don't live in a poor neighborhood.



They are specifically talking about 'over-policing', which is disproportionate focus of police activity on some sectors of the community.

Why would they call it PredPol? Don't they get that this could be just as easily interpreted as "Predatory Policing," which is quite likely more accurate?

I don't understand why this isn't obvious. Predictive analysis works if you have correct set of data. DOJ is biased towards African Americans, their data is skewed + you can't really predict what someone could do.we are humans and not rule following robots.

Unless you are very careful, machine learning is the process of projecting the biases of yesterday into the predictions of today.

The alternative to a rule driven machine is human decision makers whose emotions will play a role, sooner or later. (Prejudice and favouritism is just two emotions that are bad in this case)

With a rule driven machine, the rules and be reviewed for any direct form of prejudice and favouritism.

We've already had rule-driven frameworks: three strikes and mandatory minimums, the latter of which resulted in huge sentences for non-violent drug offenses – particularly those involving crack, a drug more prevalent in black communities, and for which the threshold to trigger a mandatory sentence was 1/100th that of powder cocaine: https://www.aclu.org/other/cracks-system-20-years-unjust-fed...

That's something I never managed to understand: if I'm a black person, and I know that current laws will unfairly publish me for drugs possession, and I know that police tenderly loves to search me at any opportunity... Doesn't it mean that I should stay as far as I can from crack and beg all my friends and relatives to do the same?

I think there comes a point where you have so little to lose and prison is so normalized in your social environment that the dignity reclaimed by defying the police state and possessing whatever you please outweighs the chance of getting caught. It's not so much a logical defect as it is a different set of premises: if prison looks like a constant threat, it stops being a deterrent and at that point you may as well have some semi-internal locus of control.

You could say that for any crime or malady of society, that it's confusing to you why millions of fellow citizens have chosen something you've found easy and obvious to avoid yourself. That doesn't address or excuse why one variation is vastly disproportionately punished.

I understand your point, but it's not about "any crime", it's about a very specific variation of a crime. Imagine DUI having a much more severe punishment and zero tolerance limits if the driver was caught between 10pm and 11pm. Now, it's 10:30pm and you've just had half a pint of beer. You need to drive back home. What do you do? Heck, you wait half an hour! Will it decrease total DUI? No. But 10pm to 11pm will be the safest time in the streets.

Poisoning your spouse now leads to a mandatory death penalty? Good. Now you'll just strangle her.

A gang is defined as 5 people or more? Excellent. All robberies are now performed by strictly no more than 4.99 people.

Crack is being cracked down upon? Fantastic. We'll smoke weed/tobacco/soap/whatever.

When there is an alternative, not using it is just dumb.

Yes, if I threaten to imprison you if any of your friends or family or neighbors commit a crime, I can press you into duty as an unfunded, undefended crimefighter. We can look to examples of the past of this strategy in action, in the glorious nations of East Germany and the USSR.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact