HU facial recognition software predicts criminality

harrisburgu.edu

2 pointsedwinjm3 comments

I'm a forensic psychiatrist. There's no plausible reason to accept the premise that criminality could ever be detected based on how someone looks... other than proliferating bias and racial profiling present in the training data.

This is a worrisome application of AI / Deep Learning.

Yes, I wonder if this is genuine or some strange attempt at pseudoscience click bait.

They do not mention any potential mechanism whereby criminality could be triggering an observable change. And it's not clear to me how a human could do this in the absence of prior knowledge either. And to round it off, they state a lack of bias but give no clear indication of how they establish that it isn't biased. Sounds more like a childish wish list ("we'll do the best research, with no mistakes")

I have an alternative algorithm to propose: 1. Input: Facial recognition and DNA sample with credit history and blood type for evaluation. 2. Output: That person is a criminal.

Literally, every single human (in the US, anyhow) has or will violate at least some law sooner or later. There are simply too many laws to avoid doing so, even to know whether one is doing so.

Reply on news.ycombinator.com