Research talk:Revision scoring as a service/Work log/2016-05-02

Monday, May 2, 2016

Can an Algorithm be Unethical?

"Can an Algorithm be Unethical? Christian Sandvig*, Kevin Hamilton Karrie Karahalios Cedric Langbort"

Summary

1. It starts with an example: HP face recognition tools suffered from a bug that couldn't locate (i.e. recognize) face of black people. It talks about how that might happen intentionally or unintentionally. For example bad training data if they were using machine learning might cause this.
2. It gives an example about photography that its standards assume human is white-skinned and does not work well with black people. Per research that they mentioned, it seems it's because of racist assumptions of fathers of photography and its standards.
3. it talks about algorithms and what is an algorithm or what is not.
4. It has a mind experiment about designing an algorithm that check surveillance footage and reports everything that needs human attention. It argues that using skin color in the algorithm is ethically wrong and if we are using machine learning skin color comes up eventually so we can do two things: 1- Either teach the classifier to ignore anything related to skin color, it argues it might not be the best option because we might consider skin color indirectly because of correlation with other features. 2- We should explicitly add predefined training data to minimize impact of skin color if it's online training, let them reduce issues by reporting false positives
5. Then it talks about ethics and what is ethics. It examples ACM's CoC and argues its content rooted in very old ethical philosophies. "virtue ethics", "consequentialism" and "deontological". It says strict "consequentialism" is a monster, justifies killing innocents for the greater good. "deontological" is just absolute rules we should respect, like copyright.
6. Then it talks how we can apply consequentialism and deontological philosophies to algorithm can do. It argues that determining consequences of an algorithm is based on circumstances thus saying an algorithm is unethical via consequentialism is not very feasible. Then talks about deontological use of ethics. It examples spam sending algorithm can be considered unethical because they mess with SMTP protocol (about headers I guess).