> Details of the algorithm and the list were initially kept secret, but when the list was finally released, it turned out that 56 per cent of Black men in the city aged between 20 to 29 featured on it.
That’s politically inconvenient, but it’s not bias. [0, 1]
It’s not politically inconvenient, it’s socially inconvenient. It’s a straightforward equation where “bad” elements in the black community need to be ruthlessly put away while not being hostile to every black person the cops come across. This seems to be something American cops are incapable of doing. The current black community isn’t helping either so it’s kind of a stalemate.
If you make it your policy to treat one group as suspicious, you'll catch a lot more people from that group. Isn't that what the police has been doing for decades?
So what kind of AI would you train with the resulting dataset? The AI doesn't know about the crimes you let slide or didn't investigate. It will reproduce whatever bias you based your policing on.
I am very skeptical. I suspect that the accuracy - metric undisclosed in the article - is dominated by “easy” cases.
For example, I can tell you where protestation / solicitation will occur next week with 100 percent positive predictive value. Same with car break ins and drug dealing.
You don’t need an algorithm for that. You can figure it out by just driving around a few neighborhoods for a couple hours.
I suspect that the algorithm does not do well on the non-obvious stuff.
Because we don't need help on the obvious stuff. Cops already know where the low level drug dealers hang out and where the prostitutes work, they don't need a computer to tell them that. Also things like 'there will probably be a shooting in this violent neighborhood sometime within the next 7 days' is completely useless information, even if correct. There is no new actionable information that can improve on the current policing situation.
It is important because otherwise it has negative utility. There is a cost associated with building, deploying, operating and maintaining it. What value is this adding if it only tells me stuff I already know?
Yeah like Plutonium Fountain on the edge of Lastarria in downtown Santiago.
I have witnessed, defended against, or interceded in 6 crimes at this one intersection. One at the fountain itself, which is technically a beautiful statue of Neptune and Amphitrite. Like urchins bathe there, dudes washing their clothes.
It's a dangerous intersection because there's five streets radiating from it. The traffic system is a little anti-pedestrian, or pedestrianophile, out to fuck pedestrians. So in Santiago there's three crosswalks per residential intersection, figuring if a pedestrian wanted to cross on the missing crosswalk all she'd have to do is cross all the other crosswalks. More efficient. So Plutonium Fountain has a really short light, just really timings, so you have to walk across on a red all the fucking time. And like motorized bicycles crossing reds, cars almost running you over...several traffic violations, just rough. It's also the point where the grid of the city gets warped by the Cerro Santa Lucía, beyond there it is all squares, this is where the Manhattan geometry meets the actual geography, the Cerro Huelén as its also called and the road to the river. Like a bottleneck.
This is where the shit hits the fan.
I can predict with 1 likelihood there'll be twenty crimes at that fountain in the next month. Not with machine learning! Just spinal. And of course I'm not going to waste the sight I have on figuring out how many more crimes there will be, but there will be twenty and there will be more. Like twenty and six, twenty and seven, there are twenty and there are also seven. Twenty and some other amount.
(OT Advice: you may want to check Lastarria, Santiago (and the Fuente Nettuno itself, at -33.4416,-70.6439) on Google Maps User Content or similar. It's unmissable.
I'd so much love to be there. How is crime against people and property in the area?
BTW: it should probably be "Poseidon's Fountain"... Pluto's part would be underground.)
Right it should be Poseidon's Fountain--in this case on the plaque it's called Neptune's Fountain--but I call it according to the next planet, Pluto (I learned there were 9 planets, that's what I learned when I was 5, Fuck the Law if the Law says there's 8, distinction is meaningless). So, why do I call it Plutonium Fountain? Because it's always shining, "shine", radiation, 5 streets radiate from there and so much crime radiates from there, in fact just talked to a cop at Plutonium Fountain an hour ago, said there were a lot of hold-ups and prostitution--in addition to the wannabe pimps, the rape extortionists, the drug dealers insulting you for not buying, the abducters who worked for the Walmart franchise--and I was like, yeah Fernando (cop's name was Fernando), glad I named it Plutonium Fountain. Always shining. Shine is a codeword for radiation, used at Los Alamos in the 40's, building the bomb, out of Plutonium in fact. Atom named after the god Pluto.
And in fact there is an underground. It's on the same small block as the fountain. There's a stairwell leading to a parking garage underneath the street which you need to give the license plate number of the car you parked there to get into--there's a shitters and urinals for men, something for women, no idea if it has any toilet paper but it's not completely disgusting.
So let me reiterate:
Plutonium Fountain.
So to answer your question about crime, crime against people and property in the area: high. Very high. Fully turgid. Well it depends on your height and your "stupid gringo" factor--mine is negative despite wearing an American flag on my leather jacket, because of my dominance and native Chilean Dialect, which shocks people when they hear it coming out of me, like a perfectly dubbed character in a Hollywood movie.
And like...yeah you'll get mugged a lot if you commute through this fountain. Get raped couple times a year, no matter what your pronouns are. Let me repeat, fellow Hacker News user, crime is high at this corner, spesh if your reading comprehension is bad, highest crime of anywhere in my life, worse than Chinatown in San Francisco in 2012, I narrowly dodged getting killed there, I narrowly dodged getting killed here.
Manhattan Geometry meets Euclidean Geometry, shit hits the fan because people go apeshit.
The only other corner like this is Alameda and Lastarria, coupla blocks away. Like, fuck, twenty police riot control vehicles on Fridays, lines of policemen, policemen in riot gear armor and clear shields, suckafucking rocks attempting brain surgery flying through the air in my trajectory, running to safety barely making it. Explosives, fireworks as explosives, police retreating and thugs advancing, All Hallow's Eve and some prostitutions visible from my nest bicycles and woafs, then I cheer the police on, and police advancing and thugs retreating.
Paramilitary, the Brigada Newén, with their presh matching white costumes black backpacks with equipment eg first aid, gas masks for tear gas, plastic masks for rubber bullets to the face, and the thing I found most endearing their round shields that said "BRIGADA NEWEN". Nonviolent, crossing Alameda in single file, they had practiced tactics, they were really cool.
Like as the graffiti nearby said, GUERRA SOCIAL. Always sum wench screaming "¡Paco Culiao!" Dude nonstop on Fridays, it's like a happy hour, always hearing "¡Paco Culiao!" there. I hate writing that but I can't describe it any other way, just the same as you have to tell people what the n-word actually is before you can refer to it as the n-word, history professors are forced to pronounce it in class. Same deal.
Social war. Class warfare. Somebody has to suffer it, and Americans had a lot to do with how things got like this, the oppressed classes in Chile lived 50 years of being treated like shit. So it might as well be me that suffers it, I'm a gringo, hell...
Wasn't it known in the 70s that something like 90% of crime occurs on a select streets? Not sure we need AI to make this prediction and achieve this accuracy.
Crime statistics aren't racist, people's explanations for "why" can be. If it's 90% accurate, then great? Do we need the algorithm to perform worse to make people feel better?
> Crime stats aren't collected in a scientific vacuum. They reflect the previous strategies the police used to fight crimes.
It may also just reflect reality. Let's say you have a group of people that commit a specific kind of crime significantly more often than another group, i.e. white male investment bankers are more likely to commit tax fraud than female black nurses. Any reasonable policy fighting this kind of crime would have to look biased against white men when it comes to enforcing tax fraud. I think that no one would reasonably call such enforcement policies bad or racist.
But when black men are significantly more likely to be prosecuted for violent crime, suddenly it's a racist policy and must be the racist polices fault, because that's the only acceptable answer.
Profiling people on what they do is very different from profiling people based on what they are. Conflating both is a perfect definition of a racist policy.
If you keep targeting the same people, and you punish their crimes harder than any other people, you will get a data set that will bias any algorithm. The AI won't know that the police stops a disproportionate number of black drivers, or that they are more likely to prosecute a black criminal than a white one. It might not understand the biases that leads to harsher sentences.
In other words, an algorithm trained on our biases will just try to reproduce our biased outcomes.
I don't disagree that american policing practices are horrible and a significant number of racist police officers will cause the data to be skewed and that would cause AI trained on that data to be skewed aswell.
What I am rejecting is that the underlying "true" data wouldn't show differences between different races. Again, tax fraud is likely commited to a much higher degree and in much higher numbers by white people (especially investment bankers). An AI wouldn't be racist to predict a higher likelyhood of a white man commiting it than a black woman. But somehow predicting a higher likelyhood of violent crime by a black male (especially if they receive wellfare) is racist because it must be so, even if the underlying data shows this to be correct.
My argument is simple, if your statistics contains racial data, either directly or transitively, than you will find statistical differences between them. It becomes a racism issue if the polices based on that data are racist, i.e. harsher sentences for the same crime etc.
The London police do race quotas in stop and search. So for every X they stop they need to step some proportional k*Y too.
No matter how you do it you will end up stopping people in "bad" neighbourhoods more. So if that is considered a problem you need to give up patrolling or do unnessecary stopping in other places to weigh up due to racial composition in the worse of neighbourhoods.
Very good points! In addition, the "AI" can be used as a kind of shield to hide bias or racist policies, because it is "blamed" for the tainted outcomes and not those who have created it or the creation of the pretense that the AI couldn't be wrong.
For petty crimes that’s true. Unfortunately it’s pretty hard to hide or fudge murders which are still committed a disproportionate amount by blacks. Usually black on black.
A long unbroken history of the cops being racist is what got us into a situation where black communities are less likely to report those crimes in the first place. Those activists may have a point.
> Chattopadhyay concedes that the data used by his model will also be biased, but says that efforts have been taken to reduce the effect of bias and the AI doesn’t identify suspects, only potential sites of crime. “It’s not Minority Report” he says
If it doesn't identify suspects then how do they know this information?
"Details of the algorithm and the list were initially kept secret, but when the list was finally released, it turned out that 56 per cent of Black men in the city aged between 20 to 29 featured on it."
> Previous efforts to use AIs to predict crime have been controversial because they can perpetuate racial bias. In recent years, Chicago Police Department has trialled an algorithm that created a list of people deemed most at risk of being involved in a shooting, either as a victim or as a perpetrator. Details of the algorithm and the list were initially kept secret, but when the list was finally released, it turned out that 56 per cent of Black men in the city aged between 20 to 29 featured on it
«Previous efforts», «In recent years, Chicago Police Department has trialled an algorithm that created a list of people». The work of Chattopadhyay is meant to reduce the bias and to avoid outputting names but just places.
- what does 90% mean, is it bad or good? That depends on the baseline and the prior distributions.
- on what dataset is the 90% computed? On past crimes? What about sampling bias in this dataset? What if the data focuses on some specific neighborhoods (oh wait...)? What about missing crimes in the data?
In ML, data should fairly well represent a phenomenon and in this case, it does not represent crime occurrence, it only represents crimes detected and documented.
The submitted divulgative article did not quite clarify what the AI is doing - it seems like weather forecasting on crime related events, but I expected more content about "how" and "how well"; for example, you do not need AI to determine "how much", and it takes some cleverness to instead predict "when", which is one of the interesting points you want explicit.
I am unfortunately lacking the time now to examine the material (hopefully I will later). For now, it seems clear that the effort is not really about crime prediction but to show that some methodology can produce unbalanced police planning:
> We introduce a stochastic inference algorithm that forecasts crime by learning spatio-temporal dependencies from event reports, with a mean area under the receiver operating characteristic curve of ~90% in Chicago for crimes predicted per week within ~1,000 ft. Such predictions enable us to study perturbations of crime patterns that suggest that the response to increased crime is biased by neighbourhood socio-economic status, draining policy resources from socio-economically disadvantaged areas, as demonstrated in eight major US cities
A.I. should not be used for law enforcement because it just creates an infinite pretext for whatever they want to do that would otherwise require probable cause, reasonable suspicion, presenting a case to a prosecutor or a judge, and all the other inconvieniences your inalieable rights create for them.
Law enforcement use cases for machine learning are nothing more than a laundering of accountability for individual decisions. The whole notion of algorithms being somehow autonomous is the most pernicious and magically anthropomorphic falsehood supposedly smart people believe, imo. I did a privacy impact assessment many years ago for a government agency who planned to outsource their benefits fraud detection to an "AI" as a service, and the service just happen to be run by a credit bureau company, effectively giving complete benefits and other personal information "for analysis" to a literal scumbag consumer credit reporting company. The project managers were trying to convince me and everyone else that somehow this data would not get re-used and integrated into the credit agencies other data sets and resold among the others - because "they're using AI, nobody there is going to see this data."
I think similar law enforcement use cases for "AI" are just as magically stupid.
36 comments
[ 2.7 ms ] story [ 77.9 ms ] threadThat’s politically inconvenient, but it’s not bias. [0, 1]
[0] https://home.chicagopolice.org/wp-content/uploads/2014/12/20...
[1] https://www.chicagotribune.com/news/criminal-justice/ct-2021...
So what kind of AI would you train with the resulting dataset? The AI doesn't know about the crimes you let slide or didn't investigate. It will reproduce whatever bias you based your policing on.
For example, I can tell you where protestation / solicitation will occur next week with 100 percent positive predictive value. Same with car break ins and drug dealing.
You don’t need an algorithm for that. You can figure it out by just driving around a few neighborhoods for a couple hours.
I suspect that the algorithm does not do well on the non-obvious stuff.
Why not handle low-hanging fruit first, and then try progressively (heh) harder situations?
See news coverage on Dutch cities using algorithms to predict financial crime/fraud. Ethnic profiling quickly becomes a thing.
Because we don't need help on the obvious stuff. Cops already know where the low level drug dealers hang out and where the prostitutes work, they don't need a computer to tell them that. Also things like 'there will probably be a shooting in this violent neighborhood sometime within the next 7 days' is completely useless information, even if correct. There is no new actionable information that can improve on the current policing situation.
I have witnessed, defended against, or interceded in 6 crimes at this one intersection. One at the fountain itself, which is technically a beautiful statue of Neptune and Amphitrite. Like urchins bathe there, dudes washing their clothes.
It's a dangerous intersection because there's five streets radiating from it. The traffic system is a little anti-pedestrian, or pedestrianophile, out to fuck pedestrians. So in Santiago there's three crosswalks per residential intersection, figuring if a pedestrian wanted to cross on the missing crosswalk all she'd have to do is cross all the other crosswalks. More efficient. So Plutonium Fountain has a really short light, just really timings, so you have to walk across on a red all the fucking time. And like motorized bicycles crossing reds, cars almost running you over...several traffic violations, just rough. It's also the point where the grid of the city gets warped by the Cerro Santa Lucía, beyond there it is all squares, this is where the Manhattan geometry meets the actual geography, the Cerro Huelén as its also called and the road to the river. Like a bottleneck.
This is where the shit hits the fan.
I can predict with 1 likelihood there'll be twenty crimes at that fountain in the next month. Not with machine learning! Just spinal. And of course I'm not going to waste the sight I have on figuring out how many more crimes there will be, but there will be twenty and there will be more. Like twenty and six, twenty and seven, there are twenty and there are also seven. Twenty and some other amount.
That is what the article hype seems to be about.
I'd so much love to be there. How is crime against people and property in the area?
BTW: it should probably be "Poseidon's Fountain"... Pluto's part would be underground.)
And in fact there is an underground. It's on the same small block as the fountain. There's a stairwell leading to a parking garage underneath the street which you need to give the license plate number of the car you parked there to get into--there's a shitters and urinals for men, something for women, no idea if it has any toilet paper but it's not completely disgusting.
So let me reiterate:
Plutonium Fountain.
So to answer your question about crime, crime against people and property in the area: high. Very high. Fully turgid. Well it depends on your height and your "stupid gringo" factor--mine is negative despite wearing an American flag on my leather jacket, because of my dominance and native Chilean Dialect, which shocks people when they hear it coming out of me, like a perfectly dubbed character in a Hollywood movie.
And like...yeah you'll get mugged a lot if you commute through this fountain. Get raped couple times a year, no matter what your pronouns are. Let me repeat, fellow Hacker News user, crime is high at this corner, spesh if your reading comprehension is bad, highest crime of anywhere in my life, worse than Chinatown in San Francisco in 2012, I narrowly dodged getting killed there, I narrowly dodged getting killed here.
Manhattan Geometry meets Euclidean Geometry, shit hits the fan because people go apeshit.
The only other corner like this is Alameda and Lastarria, coupla blocks away. Like, fuck, twenty police riot control vehicles on Fridays, lines of policemen, policemen in riot gear armor and clear shields, suckafucking rocks attempting brain surgery flying through the air in my trajectory, running to safety barely making it. Explosives, fireworks as explosives, police retreating and thugs advancing, All Hallow's Eve and some prostitutions visible from my nest bicycles and woafs, then I cheer the police on, and police advancing and thugs retreating.
Paramilitary, the Brigada Newén, with their presh matching white costumes black backpacks with equipment eg first aid, gas masks for tear gas, plastic masks for rubber bullets to the face, and the thing I found most endearing their round shields that said "BRIGADA NEWEN". Nonviolent, crossing Alameda in single file, they had practiced tactics, they were really cool.
Like as the graffiti nearby said, GUERRA SOCIAL. Always sum wench screaming "¡Paco Culiao!" Dude nonstop on Fridays, it's like a happy hour, always hearing "¡Paco Culiao!" there. I hate writing that but I can't describe it any other way, just the same as you have to tell people what the n-word actually is before you can refer to it as the n-word, history professors are forced to pronounce it in class. Same deal.
Social war. Class warfare. Somebody has to suffer it, and Americans had a lot to do with how things got like this, the oppressed classes in Chile lived 50 years of being treated like shit. So it might as well be me that suffers it, I'm a gringo, hell...
photosphere photo of the statue and intersection
Crime statistics aren't racist, people's explanations for "why" can be. If it's 90% accurate, then great? Do we need the algorithm to perform worse to make people feel better?
Say for example that a white guy smoking a joint gets a slap on the wrist, and the black guy gets the book thrown at him.
What would an AI derive from what it's given?
Crime stats aren't collected in a scientific vacuum. They reflect the previous strategies the police used to fight crimes.
It may also just reflect reality. Let's say you have a group of people that commit a specific kind of crime significantly more often than another group, i.e. white male investment bankers are more likely to commit tax fraud than female black nurses. Any reasonable policy fighting this kind of crime would have to look biased against white men when it comes to enforcing tax fraud. I think that no one would reasonably call such enforcement policies bad or racist.
But when black men are significantly more likely to be prosecuted for violent crime, suddenly it's a racist policy and must be the racist polices fault, because that's the only acceptable answer.
If you keep targeting the same people, and you punish their crimes harder than any other people, you will get a data set that will bias any algorithm. The AI won't know that the police stops a disproportionate number of black drivers, or that they are more likely to prosecute a black criminal than a white one. It might not understand the biases that leads to harsher sentences.
In other words, an algorithm trained on our biases will just try to reproduce our biased outcomes.
What I am rejecting is that the underlying "true" data wouldn't show differences between different races. Again, tax fraud is likely commited to a much higher degree and in much higher numbers by white people (especially investment bankers). An AI wouldn't be racist to predict a higher likelyhood of a white man commiting it than a black woman. But somehow predicting a higher likelyhood of violent crime by a black male (especially if they receive wellfare) is racist because it must be so, even if the underlying data shows this to be correct.
My argument is simple, if your statistics contains racial data, either directly or transitively, than you will find statistical differences between them. It becomes a racism issue if the polices based on that data are racist, i.e. harsher sentences for the same crime etc.
Nobody claimed this
The London police do race quotas in stop and search. So for every X they stop they need to step some proportional k*Y too.
No matter how you do it you will end up stopping people in "bad" neighbourhoods more. So if that is considered a problem you need to give up patrolling or do unnessecary stopping in other places to weigh up due to racial composition in the worse of neighbourhoods.
> Chattopadhyay concedes that the data used by his model will also be biased, but says that efforts have been taken to reduce the effect of bias and the AI doesn’t identify suspects, only potential sites of crime. “It’s not Minority Report” he says
"Details of the algorithm and the list were initially kept secret, but when the list was finally released, it turned out that 56 per cent of Black men in the city aged between 20 to 29 featured on it."
So it is Minority Report... Literally.
> Previous efforts to use AIs to predict crime have been controversial because they can perpetuate racial bias. In recent years, Chicago Police Department has trialled an algorithm that created a list of people deemed most at risk of being involved in a shooting, either as a victim or as a perpetrator. Details of the algorithm and the list were initially kept secret, but when the list was finally released, it turned out that 56 per cent of Black men in the city aged between 20 to 29 featured on it
«Previous efforts», «In recent years, Chicago Police Department has trialled an algorithm that created a list of people». The work of Chattopadhyay is meant to reduce the bias and to avoid outputting names but just places.
RTFA...
- what is the metric exactly?
- what does 90% mean, is it bad or good? That depends on the baseline and the prior distributions.
- on what dataset is the 90% computed? On past crimes? What about sampling bias in this dataset? What if the data focuses on some specific neighborhoods (oh wait...)? What about missing crimes in the data?
In ML, data should fairly well represent a phenomenon and in this case, it does not represent crime occurrence, it only represents crimes detected and documented.
So I found another divulgative article at Phys.org : https://phys.org/news/2022-06-algorithm-crime-week-advance-r...
and a news piece / press release at UChicago.edu : https://biologicalsciences.uchicago.edu/news/algorithm-predi...
and finally, the original research article, at: https://www.nature.com/articles/s41562-022-01372-0
I am unfortunately lacking the time now to examine the material (hopefully I will later). For now, it seems clear that the effort is not really about crime prediction but to show that some methodology can produce unbalanced police planning:
> We introduce a stochastic inference algorithm that forecasts crime by learning spatio-temporal dependencies from event reports, with a mean area under the receiver operating characteristic curve of ~90% in Chicago for crimes predicted per week within ~1,000 ft. Such predictions enable us to study perturbations of crime patterns that suggest that the response to increased crime is biased by neighbourhood socio-economic status, draining policy resources from socio-economically disadvantaged areas, as demonstrated in eight major US cities
Law enforcement use cases for machine learning are nothing more than a laundering of accountability for individual decisions. The whole notion of algorithms being somehow autonomous is the most pernicious and magically anthropomorphic falsehood supposedly smart people believe, imo. I did a privacy impact assessment many years ago for a government agency who planned to outsource their benefits fraud detection to an "AI" as a service, and the service just happen to be run by a credit bureau company, effectively giving complete benefits and other personal information "for analysis" to a literal scumbag consumer credit reporting company. The project managers were trying to convince me and everyone else that somehow this data would not get re-used and integrated into the credit agencies other data sets and resold among the others - because "they're using AI, nobody there is going to see this data."
I think similar law enforcement use cases for "AI" are just as magically stupid.