The algorithm police is coming. Will it have teeth? | Hacker News Reader

The algorithm police is coming. Will it have teeth?

algorithmwatch.org

12 pointsn_kb5 comments

Why do I have a feeling that the way these laws will work will be to add some "academicaly derived, politcaly approved" bias to certain machine learning models to push the results in the "correct" direction?

That's an unnecessarily cyncial and contrarian view. It's 2020, regulating software and data is a thing (copyright, HIPAA, GDPR, etc). It's about time we get on to regulating software design. We wouldn't regulate drivers and not regulate car manufacturers.

And yet, we also wouldn’t demand car manufacturers change how they build a car because races, sexes, or genders differentially purchase, say, lexuses vs hondas.

I agree with this. I saw an article awhile back detailing issues with bank loan algorithms. Although race/gender weren't a determination, it looked at credit scores and ability to pay the lone back.

The majority of minorities fell into this category and the algorithm was deemed 'racist'.

Fixing the algorithm will not solve the problem and more likely than not, it will make the loan system unstable for everyone by giving out loans to people that can't afford them and most likely will never pay them back.

While there's not much meat to the article, the topic of whether and how to police algorithmic decision-making is an interesting one.

On the one hand, biased resource allocation can be insidious and giving a free pass to any discrimination that's implemented within a black box labeled "algorithm" seems foolish.

On the other hand, none of the alternatives I see:

* banning a predefined list of analytical practices, * outlawing a list of input types that may be used, * second-guessing the quality of not-transparently-unjust logic to assign negligence, or * defining a "fair" outcome that's not to be deviated from

seem both administrable and sufficient, especially as the stakes go up and complexity necessary to improve decision outcomes becomes more complex (turn left-right-straight given a LIDAR image, predict likelihood to reject an organ given patient history and donor/patient DNA profiles).

Reply on news.ycombinator.com