In chapter 6 of the book Weapons of Math Destruction the author discusses about how some companies use WMD based tests to screen applicants unfairly. At first, we see an example of Kronos which is a personality screening test. Because of Kronos, Kyle was rejected from a minimum wage job since it detected personality disorders from Kyle’s answers. Such systems create a negative feedback loop by rejecting candidates with mental health history since a rejection due to mental health reasons is likely to cause the candidate’s conditions to exacerbate.
At the same time, the author compares these human resources filtering tools to the ones used in sports. In the case of sports, data on athletes are collected similarly. However, since every athlete gets individual attention by the models, these models are required to be updated every now and then based off the performance updates of these athletes. As a result, because of this positive feedback, the model improves and becomes fair. However, in the case of human resource WMDs, these are used to filter out thousands of candidates in a cheap way. It is intended to make the hiring process easier. As a result, without any extraordinary circumstances these models are usually never updated. So, due to a lack of update, the model becomes incompetent and misses out on potential star employees.
In the name of science and improvement, some models have been updated to keep track of a candidate’s entire virtual existence on social media and other websites to see whether a candidate eventually becomes an good hire or not. But, these models tend to intrude the candidate’s private life. In fact, because of possible noisy data, good candidates can get filtered out for no good reason.
Finally, I think that the fundamental problem with a recruitment WMD or any other WMD for that matter is a lack good mix of good data. As a model is only as good as its data, if the data is biased, the model will only reflect the biased data from a stereotypical human society. Sometimes, lack of good data encourages the designers of such models to adopt proxy data which can create further misleading models. So, in my opinion, it is in our best interest to make the most out of the scale and robustness that WMDs give us. But, we should be aware of the data that we are training the WMDs on. At the same time, we should also ensure that we have a positive feedback loop in place to help the model get better with updates.
