The application of Artificial Intelligence in human resources has an enormous potential, but it faces some challenges that are becoming very evident.
To understand the application in the field of human resources, it is useful to explore the difference between supervised and unsupervised ML. The following article by Andrew Ng can help you understand this distinction What Artificial Intelligence Can and Can’t Do Right Now. ML is called supervised, when inputs are mapped to outputs. For example, you have some photographs and the machine learns to recognize if there are human faces in them. In this context, the key is that you need many examples of photos, with and without faces, to be able to teach the machine. Once the machine learns, it can differentiate. This method is used to classify and study regressions.
Against this background, the applications to human resources are multiple. A good example of supervised ML is CV screening based on certain predetermined parameters. There are many similar applications. The application of ML becomes more problematic in unsupervised or more sophisticated environments. For example, the theory says that machines evaluate CVs objectively, while humans can value some attributes and ignore others depending on their own background and preferences. However, if the data with which we are training the machines has bias, we will be introducing biases unconsciously, and potentially at a larger scale.
First of all, we must take into account that when talking about people, data is sensitive, difficult to obtain and always subject to the data protection law applicable in each country. Secondly, in the field of people management, you need to explain the decisions you make with transparency. Additionally, the impact of an error in anomalous cases (I mean a case that does not respond to the usual correlations) is not acceptable. Hiring a person is not the same as recommending a song.
In the case of Amazon, as reported in A Practical Approach to Detecting and Correcting Bias in AI Systems, they built their CV database with CVs submitted during the last 10 years, most of which were men. Without a more representative sample, the algorithm developed a bias against women. Unfortunately for ML, the selection of people is an extremely complicated issue. First, it could be said that human resources professionals also learn following a ML methodology. We learn and make predictions based on our past experience. For this reason, a rich and diverse experience often helps people to have less bias. Of course, years of work and having conducted many interviews favor a broader vision. The value of experience is fundamental in this field. Additionally, HR professionals must be open and have mental flexibility to detect new trends and adapt their criteria to avoid hiring for the future based only on skills that have been successful in the past. Josh Bersin's reflection on this topic is interesting in the article Hire Leaders for What They Can Do, Not What They Have Done (including an interesesting mention to the Peter principle).
In summary, the application of ML to the screening of candidates has four major drawbacks: 1) Difficulty of obtaining a sample without bias. 2) The risk of error in sensitive issues. 3) Intrinsically, ML is based on learning from the past, as it is trained with existing data, and presents difficulties in anticipating future needs. 4) Transparency and understanding of the criteria are necessary, so the use of unsupervised techniques is limited.
The above mentioned difficulties limit, for the moment, the applications of ML in the most sensitive and complex human resources activities. In any case, the potential of AI in this field is undoubtedly enormous.