Reports suggest that an algorithm that was being tested as a recruitment tool by online giant Amazon was sexist and had to be scrapped. Whilst unconfirmed, this news raises a number of questions about how artificial intelligence or machine learning should be applied to talent acquisition in a way that will not pick up the unconscious bias of humans.
Top employers continue to receive huge numbers of applications for each open position and ultimately, the majority of applicants--upwards of 95 percent--are turned down. With AI, employers should have an even greater ability to quickly flag candidates that have certain key indicators of success to streamline the selection process and nurture this talent ahead of competitors.
It is important though that a tool never does this based on historical data alone. Efforts must be made to constantly improve the robustness of any tool to help leading employers benefit from the best possible early evaluation of applicants based on responses given within online application forms. This can be achieved by collecting enormous amounts of structured and unstructured data, processes that data using the best from thousands of machine-learning algorithms to most accurately predict outcomes, and refines that process as it learns.
Constant machine learning will work to reduce unconscious biases and enhance diversity by uncovering strong candidates who may have gone unnoticed in a non-intelligent or manual process. In turn, recruiters gain insight and reasoning into which characteristics score the strongest.
Oleeo commissioned the Department of Computer Science at University College London to look into how algorithms can ensure that they do not inadvertently fall into gender bias, as Amazon appears to have done.
It revealed that removing any wording or phrases that could unconsciously predict the gender of a candidate would enable algorithms to make any gender prediction to be no better than random with no direct impact from the loss of information in the transformation and de-biasing steps. In fact, more consistent disparate impact scores of close to 1.0 (i.e. no disparate impact observed) are recorded in hiring predictions undertaken in this way providing better hired prediction performance. It is also shown to have consistent negligible disparate impact across a range of hiring values, providing room for adjustment in recruitment screening thresholds without increasing disparate impact.
Working in this way allows...