Comment by jcgrillo
1 month ago
If I were in law enforcement, had no morals, and just wanted to convict as many people as possible I'd build a system that automatically assembles a virtual dossier on everyone using these data streams. Then I'd implement detection heuristics that look for interesting dossiers. These could be used as the "classified" component of a case built by parallel construction[1].
Not even. It's worse. They aren't even useful for that.
They've tried that approach but it's actually less efficient than "good old fashioned police work" because it turns out that 99/100 of your hits are gonna be lawful weirdos, 1/100 is gonna be a petty drug dealer and the career advancing prosecution you actually wanted would have been much easier to find by using normal methods like inferring that a dealer has a supplier, a spy has a handler, etc, etc and trying to suss out who those people are. The NSA figured all this out post 9/11 when they were building data haystacks in search of terrorists.
What the data haystacks do get used for is dragnet policing wherein an agency picks some crime they're gonna go hard on, pulls up a bunch of results of people who probably did it, tosses all the people who are likely to pose any risk to them (e.g. you don't see the ATF knocking on doors asking about Temu glock switches in bad parts of Detroit) and kicks in the doors of whoever's left.
The data haystacks are also really useful for witch hunts when they get egg on their face and need to make someone pay, like that time they prosecuted anyone and everyone who they could construe as having done anything to help the kid who bombed the Boston Marathon, and the January 6 people of whom a great number were certainly just hapless.
And this is in addition to the usual "opposition research" like the FBI bugging MLK and all that sort of crap.
If you had a location that was a known drug hot spot, you could use this data to see who frequented that location. Using that info, you could use "good old fashioned police work" to contact each person and get them to roll on someone else. That's much easier than sitting in a stakeout trying to ID those that come and go.
Or you watch them, find out where the stash house is, and call in an "anonymous tip" to another agency. They get a warrant, raid the stash, and it's all above board (or near enough).
Parallel construction makes the mere existence of these data sets extremely dangerous.
4 replies →
Any references to back up the suggestion that a data driven approach doesn’t work?
Not being skeptical, but curious
Surely ML is better than that? Otherwise no targeted advertising would work.
Last I checked, no one has any evidence that targeted advertising does work.
For example: https://marketinginsidergroup.com/marketing-strategy/digital...
You can do things far more interesting than that with the dossiers on everyone that absolutely exist right now and that algorithms are constantly being run over. You can frame people for crimes for which you know they will have no defense, exactly like the Stasi did, and privately confront them about it. As they plead their innocence, tell them that you want to believe them, and if they can do a little work for you, they'll not only be arrested, but be rewarded! How would you like a job at Mother Jones, or the Guardian?