4 ms·

The algorithm police is coming. Will it have teeth?

12 points by n_kb 7y ago

OneGuy123 7y ago
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?
- subjectsigma 7y ago
  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.
  - kian 7y ago
    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.
- planetzero 7y ago
  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.
perpetualpatzer 7y ago
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).

12 points by n_kb 7y ago

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?

subjectsigma 7y ago
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.
- kian 7y ago
  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.
planetzero 7y ago
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.

perpetualpatzer 7y ago

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).