AI can predict the future criminals based on facial features | Hacker News Reader

AI can predict the future criminals based on facial features

theintercept.com

27 pointschang230118 comments

I can also look at the 6 pictures and accurately guess who the bad guys are. The 3 who are frowning like they're being photographed against their will for a mug shot are bad. The 3 guys earnestly smiling and wearing suits are good.

I dunno if the sample from that page is representative of their test set, but if so, it's pretty weak.

Don't get distracted with the whole pseudoscience/phrenology angle.

The police state doesn't really care if people are criminals or not. It just wants to know who is more likely to give them trouble: A common thug who scares your voters is a problem. A white-collar criminal who makes regular campaign contributions is not.

Since that could probably be gleaned from facial characteristics (expression, grooming, style, oppressed-ethnicity-traits, etc.), I'm sure this scratches an itch.

They'll be running this on us in the US and Britain before you know it.

In US it is already working as intended - it is called "crime of doing X while black" - no need for subtle facial characteristics, recognition is easy and fool-proof.

Do you see now why it is stupid?

In the US, duh. In other countries, where the difference between the oppressor and the oppressed is more subtle, this provides "value" (like different castes of Indians, or upperclass/lowerclass/Irish/Scottish/English/Welsh in Britain).

This post by Yonatan Zunger (Head of Infrastructure for the Google Assistant, and a great thinker) on Google+ is insightful: https://plus.google.com/+YonatanZunger/posts/J8qrWvmhwNF?sfc...

The conclusion says it best:

"Contrary to the Intercept article, this paper isn't a failure of ML ethics; it's a failure of ordinary academic ethics, dressed up as machine learning. The paper is bullshit, plain and simple."

I find it interesting how hard the author focuses on disparaging the authors of this paper for daring to investigate this question, rather than trying to argue with the empirical results of their research.

If you paid attention on history lessons you'd know why going this path is very dangerous.

If you read the article, these aren't even convicted criminals -- they're just suspects from some Chinese database. This is a truly horrible piece of "research".

Things like this can only be endorsed by malicious or very stupid people (if you are getting excited about this and have Minority report pop into your head -- you're of the latter).

This is total bullshit, the best crooks all look like bank managers for a good reason

Recent discussion of this study: https://news.ycombinator.com/item?id=12983827

I'm not sure we want computers perfecting human biases. I will say that Hollywood (and probably other movie cultures as well) is not above using mean-looking people to represent heavies in film.

Which, despite its unfairness, does give us gems such as this -- https://www.youtube.com/watch?v=A8hjFmPCSjg

There is no independent data set used to validate the model(s). We have no idea how the models will generalize beyond this data set.

The big questions which need further research:

0. Beyond socio/psychopathy, are there genetic markers for criminal tendencies?

1. Do unconscious biases of other people motivate certain looking people to engage in crime?

Who chose the 1,856 source images?

Yet another example of shitty science.

Are we bringing back frenology now?

Eugenics.

Reply on news.ycombinator.com