AI: The Good, the Bad and the Biased
Artificial Intelligence is getting into our everyday life and is expected to generate significant benefits for businesses and individuals in the near future. According to Accenture, AI technologies are expected to increase productivity by 40% and double economic growth in developed economies by 2035.
Still, despite the shiny perspectives there is a “BUT”. With more and more decisions relying on AI people started to realize that blind application of algorithms can lead to disturbing biases. An algorithm was trained on Google News articles to create word embeddings — vector representations of words with which you can relate words one to another, e.g. “man is to king as woman is to queen”. When you asked the algorithm “man is to computer programmer as woman is to?”, the answer was “homemaker”[1]. Definitely not the answer everyone expects today. While sentiment analysis and the likes might not seem to be a big issue the list goes on. COMPASS one of the risk-assessment algorithmsbeing used in the US justice system to identify potential re-offenders was repeatedly accused of being biased against minorities.
The negative consequences of biased predictions are clear: they create unhealthy environment for decision making and they may lead to “self-fulfilling forecasts” (if you lock down a person wrongly identified as a potential re-offender, this person obviously will not commit a crime, creating a positive feedback loop for wrong forecasts in the future).
Despite the gloomy outlook it appears that AI is not the one to blame. Even the best algorithms are trained on the data that exists in the world and is often generated by humans in one way or another. And those humans are well-known for stereotypes and opinion biases. One research identified that first-time borrowers assigned to officers of the other gender pay, on average, 35 basis points higher interest rates compared to borrowers assigned to same-gender officers[2]. Another study demonstrated that hiring discrimination against Blacks hasn’t changed in the last 25 years[3]. Gender biases, age discrimanation, you name it.
Luckily for everyone algorithms can be fixed to eliminate bias.
But who will fix the bias in our heads?
[1] Man is to Computer Programmer as Woman is to Homemaker? Debiasing Word Embeddings. Tolga Bolukbasi, Kai-Wei Chang, James Zou, Venkatesh Saligrama, Adam Kalai
[2] Beck, Thorsten, Patrick Behr, and Andreas Madestam (2012), “Sex and Credit: Is There a Gender Bias in Lending?”, CentER Discussion Paper, 2012–062.
[3]1. Meta-analysis of field experiments shows no change in racial discrimination in hiring over time. Lincoln Quillian, Devah Pager, Ole Hexel and Arnfinn H. Midtbøen
