In fact, 89% of ShotSpotter deployments turn up no evidence of gun-related crime: [1]
And that's case reports - not necessarily detection true positives. If someone fires a gun at the sky and picks up the shell casings and drives away, with no witnesses or cameras and an inconclusive/indiscernible audio recording, police likely won't be able to file a report.
Also, I'm guessing there are some confidence indicators for each alert: those case report rates probably don't capture the full story even when accounting for hypothetical inconclusive true positives. Police are likely (and should be) obligated to investigate every detection, including every detection rated as low-confidence.
Plus that's just for Chicago. I doubt they actually achieve their claimed 97% accuracy anywhere, but it could be higher in other areas. Especially areas where loud noises are less common.
The main issue seems to be when it's used as evidence. I think it shouldn't ever be admissible as evidence of anything. And to mitigate some of the other issues that site raises, it'd be nice if there were some kind of "no fishing expeditions" law, where police responding to a gunfire alert must exclusively look for evidence of a fired gun in that vicinity and not use it as an excuse to scavenge for low-hanging fruit by harassing / arresting people in the area for unrelated matters.
I think the least wasteful approach would be to build rapport with the community and rely on their reports. Probably not an easy thing, but the most valuable things rarely are.
Shotspotter is a very useful tool IMO.
This generally doesn't happen. It originated as a slander against the people who lived around where Kitty Genovese was killed, as an after-the-fact excuse for the lack of police response.
> While there was no question that the attack occurred, and that some neighbors ignored cries for help, the portrayal of 38 witnesses as fully aware and unresponsive was erroneous. The article grossly exaggerated the number of witnesses and what they had perceived. None saw the attack in its entirety. Only a few had glimpsed parts of it, or recognized the cries for help. Many thought they had heard lovers or drunks quarreling. There were two attacks, not three. And afterward, two people did call the police. A 70-year-old woman ventured out and cradled the dying victim in her arms until they arrived. Ms. Genovese died on the way to a hospital.
https://en.wikipedia.org/wiki/Murder_of_Kitty_Genovese#Accur...
For instance, the murder of Seth Rich in my home city (ignore the stupid conspiracy theories) was not followed by a call but was found by shotspotter which is how it was responded to.
Also, that's still two out of 38 people who called. If only 36 people were there, without those two, it's possible no one might have called.
This tends to be the issue with a lot of "AI" technology tied to police work. It does work to some extent, but it mostly tells you things you already know.
That seems incompatible with roughly nine out of ten shotspotter reports ultimately being false positives.
1) False positives lead to under reaction. This is the "boy who cried wolf" issue.
2) Extra false positives put strains on resources, as they produce spurious false reports that must be dealt with in parallel with other requests.
Combine these two, and I'm dubious that shotspotter would actually make the police more responsive to real shootings.
Alert fatigue is a big problem for any alerting system, but here the baseline is 911 reports. If they receive a 911 report for a shooting, they're going to arrive after the report. If they receive a ShotSpotter alert before a 911 report for the same incident, then they're probably going to arrive as soon or sooner than they would have otherwise. If they receive an alert but don't receive a 911 report, then any response time will be faster than no response.
Not sure if these are backed by data, but there are many quotes of law enforcement saying response times are decreased. I could paste a bunch of quotes but I know it doesn't really mean much.
This study used "gunshot victim prehospital time" as a proxy; arguably a more important proxy:
>Use of ShotSpotter detection technology decreases prehospital time for patients sustaining gunshot wounds - https://pubmed.ncbi.nlm.nih.gov/31425474/
Some other studies indeed found it did not seem to decrease response time, though.
Another factor is that it tries to precisely locate the source of the shot. It's hard for humans to infer the location and direction of a supersonic bullet's source if they're going just by hearing and didn't see the incident, while in theory an automated system could determine it more accurately. I don't know how helpful this is empirically, though.
So you may be right. I think most of the value probably comes from cases where they otherwise wouldn't have received a 911 report, or only would have received one much later.
This seems true, and if it happens a lot that would imply that the system is worthwhile.
> If they receive an alert but don't receive a 911 report, then any response time will be faster than no response.
Assuming it’s real, yes. If it’s not, it’s taken a cop on a wild goose chase when they could have been doing something else.
> Not sure if these are backed by data, but there are many quotes of law enforcement saying response times are decreased. I could paste a bunch of quotes but I know it doesn't really mean much.
Some of these are certainly true and in good faith, but as a general rule I don’t trust quotes from cops about whether or not they need new powers or tools since their answer is always “yes”. Sometimes they’re right, of course, but we need to evaluate more unbiased sources of data to correctly determine what tools are worth it.
It’s a bit like asking me if I need an M1 to do my job better. I might be incidentally right, but I also have extra motivations too.
And that’s before we talk about tools that are more ambiguous. Remember that this data will be used to prosecute people too, if it’s not accurate then the consequences might be worse than just sending out on pointless calls.
> This study used "gunshot victim prehospital time" as a proxy; arguably a more important proxy:
Yeah, that’s compelling.
> Another factor is that it tries to precisely locate the source of the shot. It's hard for humans to infer the location and direction of a supersonic bullet's source if they're going just by hearing and didn't see the incident, while in theory an automated system could determine it more accurately. I don't know how helpful this is empirically, though.
If it works, I agree. The article implies that it is not accurate in determining the location of a gun shot. The extremely high false positive rate also could be from the system accurately detecting a gunshot and inaccurately placing its location.
At a cost of $33 mil a year for the city of Chicago, I personally can’t help but wonder if there’s something else that could be done with that money to reduce the crime rate.
The supersonic crack is a factor, but it’s usually not most of the sound. Typically you will not be able tell a difference between subsonic and supersonic ammunition in the same caliber, it’s all too loud to notice. It’s only once it’s suppressed that you can tell.
Many common handgun calibers are naturally subsonic too, such as the fairly popular .45ACP, as well as heavier 9mm loads. The march 22lr ammo I use is naturally subsonic (faster, trans sonic ammo is less accurate), and it’ll still destroy your hearing without ear protection on.
Maybe any fast, loud noise is difficult to place and the supersonic crack is pretty irrelevant, though? Or maybe sounds tend to just be hard to place in general, similar to how eye witness reports tend to not be very reliable?
Even if one assumes all the other parts are unreliable, I think the most important question is what percentage of confirmed gunfire was identified by the system which otherwise would never have been discovered, or only discovered more than 30 minutes later.
If there are even a few cases like that, then I think that could lead to someone potentially being saved when they'd otherwise face certain death; potentially due to bleeding out over a long period of time with no one nearby to help. To me, that alone would justify the cost.
This other HN comment is one example:
>For instance, the murder of Seth Rich in my home city (ignore the stupid conspiracy theories) was not followed by a call but was found by shotspotter which is how it was responded to.
Given enough instances like that, it seems pretty plausible that in at least one case, someone who would've died could be rescued.
I'd like to suggest that it might lead to slower response times in the cases of false negatives: someone calls in a report and the ShotSpotter doesn't make an alert. If ShotSpotter tends to be over sensitive police might assume if they don't get an alert from it something definitely didn't occur and take their time
A system could be considered generally "sensitive" in some way (e.g. seemingly firing in response to any loud noise) while actually performing poorly in terms of statistical sensitivity (the percentage of actual gunshots that the system detects and alerts on with high confidence).
The article discusses instances where it is used in trial and for that it is obviously pretty weak - surprised it can be used at all.
If ShotSpotter is limited to registering the the sound of gunshots, then it is limited to firearm-discharge-related crime which is a subset of gun-related crime.
If you track the number of calls to 911 that are "dead-ends" because there are no arrests I'm sure it would be high too.
It also doesnt say the error rate for when the police aren't called. If criminals know that if they shoot a gun there is a 50% chance a police car will drive through 5 minutes later - wouldn't that deter you?
In the end though every game had to be translated to the controller and there was always enough of a change to make it feel unintuitive. If Valve had paid an intern to go through the top games each day and make official layouts for games that roughly lined up with each other I'd probably be using one now.