Police Use AI to Flag 'Suspicious' Driving, and Hacker News Readers Call It a Surveillance Dragnet
원문을 읽지 못해 해커뉴스 토론으로 요약했습니다
- A commenter quotes the article's lead example: a woman was pulled over for spending too much time in the left lane, which the commenter calls stupid but targeted rather than a dragnet.
- Commenters point to ACLU reporting on AI-driven policing (aclu.org/news/privacy-technology) as the background for the story, and one calls for a federal ban on owning or selling access to license plate reader networks.
- One thread calls the automated alerts dragnets illegal under the Fourth Amendment, while a reply argues plate readers are the opposite of a dragnet: police have always been able to watch cars and read plates, and only indiscriminate stop and search makes a dragnet.
- Commenters say racial profiling by such models would be nearly impossible to prove because the system usually cannot explain why it flagged someone, and one suggests feeding counterfactual versions of the same image, with the driver edited to a different race, to see if the output changes.
- A commenter who builds clinical software for doctors and nurses describes a data science team testing models for known biases, such as the false belief that Black patients have higher pain tolerance, and blocking any model that reproduces them from production.
Hacker News 의견들
Why was this FBI guy so forthcoming with a reporter about his agency's arguably civil rights abuses? Good that he was, I guess, but that's not how feds usually act in my experience.
He's not confessing, he's bragging. Cops are obsessed with anything that lets them surveil people, and they'll buy whatever unnecessary toy tickles them, right down to surplus MRAPs.
Ellison said citizens will be on their best behavior because everything is recorded and reported. That falls apart when people who follow the rules perfectly get scrutinized anyway. Why bother following them?
The ACLU has been covering exactly this: aclu.org/news/privacy-technology/ai-driven-policing. Worth reading before the comments fill up.
How is a police officer's judgment better than the model's? The human in the loop is what makes it subjective, and I've had plenty of encounters where an officer's discretion helped me and my friends.
Reasonable suspicion isn't probable cause, and the reasonable part is absurdly subjective here. These detentions aren't legal, whatever the vehicle code says.
That's sovcit nonsense. Are you writing this from jail?
We used to call these dragnets, illegal under the Fourth Amendment, but the Constitution only seems to apply to people who can afford an attorney.
That's not what a dragnet is. Cops have always been able to watch cars and read plates. A dragnet is indiscriminate stop and search with no specific suspicion, and the left lane example is stupid but targeted.
I'm used to AI agents doing what I tell them. It's going to be rough the first time I realize one has direct authority over me.
Can AI fully legalize racial profiling? Crime correlates with age, gender, skin color, and nobody can ask the model why it flagged someone, so nobody could ever prove it.
You can feed it counterfactuals. Edit the person in the image to be white, push it back through the model, and see if its opinion changes.
I work on software doctors and nurses use, and our data scientists test for exactly this. The belief that Black patients have higher pain tolerance is false but still shows up in notes, so they try to surface those biases in the model and block it from production.