• 
• 27 min
• Cornell Tech
• 2019

Solon Barocas discusses his relatively new course on ethics in data science, following a larger trend of developing ethical sensibility in this field. He shares ideas of spreading out lessons across courses, promoting dialogue, and making sure we are really analyzing problems while learning to stand up for the right thing. Offers a case study of technological ethical sensibilities through questions raised by predictive policing algorithms.

• Computing Ethics Issues
• Algorithmic Bias
• Machine Learning

• 
• 28 min
• Cornell Tech
• 2019

Pre-trial risk assessment is part of an attempted answer to mass incarceration. Data sometimes answers a different question than the ones we’re trying to answer (data based on riskiness before incarceration, not how dangerous they are later). Essentially, technologies and algorithms which end up in contexts of social power differentials can often be abused to further cause injustice against people accused of a crime, for example. Numbers are not neutral and can even be a “moral anesthetic,” especially if the sampled data has confounding variables that collectors ignore. Engineers designing technology do not always envisage ethical questions when making decisions that ought to be political.

• Civil Surveillance
• Algorithmic Bias
• Machine Learning

• 
• 10 min
• The New Yorker
• 2019

Great breakdown of the concerns that come with automating the world without understanding why it works. Provides the principal concerns with the “hidden layer” of artificial neural networks, and how the lack of human understanding of some AI decision making makes these machines susceptible to manipulation.

• Limitations of Digital Technologies
• Scientific Innovation
• Machine Learning
• Data Mining

• 
• 27 min
• Cornell Tech
• 2019

Podcast about worker quantification in factors such as hiring, productivity and more. Dives into the discussion on why we should attempt a fair making of algorithms. Warns specifically about how algorithms can find “proxy variables” to approximate for cultural fits like race or gender even when the algorithms is supposedly controlled for these factors.

• Algorithmic Bias
• Technology and Race
• Machine Learning

• 
• 7 min
• TED
• 2017

Predictive policing software such as PredPol may claim to be objective through mathematical, “colorblind” analyses of geographical crime areas, yet this supposed objectivity is not free of human bias and is in fact used as a justification for the further targeting of oppressed groups, such as poor communities or racial and ethnic minorities. Further, the balance between fairness and efficacy in the justice system must be considered, since algorithms tend more toward the latter than the former.

• Predictive Policing
• Algorithmic Bias
• Machine Learning

• 
• 5 min
• GIS Lounge
• 2019

GIS, a relatively new form of computational analysis, can often contain algorithms with biases based on biases present in the training data from open data sources, with this case study focusing on the tendency of power-line identification data being centered around the Western world. This problem can be improved by approaching data collection with more intentionality, either broadening the pool of collected geographic data or inputting artificial images to help the tool recognize a greater number of circumstances and thus become more accurate.

• Global Impact of Technology
• Algorithmic Bias
• Machine Learning
• Object Recognition