Borocas’s research has found that machine learning in hiring, much like its use in facial recognition, can result in unintentional discrimination. Algorithms can carry the implicit biases of those who programmed them. Or they can be skewed to favour certain qualities and skills that are overwhelmingly exhibited among a given data set. “If the examples you’re using to train the system fail to include certain types of people, then the model you develop might be really bad at assessing those people,” Borocas explained.
Not all algorithms are created equal — and there is disagreement among the AI community about which algorithms have the potential to make the hiring process more fair.
One type of machine learning relies on programmers to decide which qualities should be prioritised when looking at candidates. These “supervised” algorithms can be directed to scan for individuals who went to Ivy League universities or who exhibit certain qualities, such as extroversion.
“Unsupervised” algorithms determine on their own which data to prioritise. The machine makes its own inferences based on existing employees’ qualities and skills to determine those needed by future employees. If that sample includes only a homogeneous group of people, it will not learn how to hire different types of individuals — even if they might do well in the job.
Companies can take measures to mitigate these forms of programmed bias. Pymetrics, an AI hiring startup, has programmers audit its algorithm to see if it is giving preference to any gender or ethnic group. Software that heavily considers postal codes, which strongly correlates with race, is likely to have a bias against black candidates, for example. An audit can catch these prejudices and allow programmers to correct them.
Stella IO also has humans monitoring the quality of the AI. “While no algorithm is ever guaranteed to be foolproof, I believe it is vastly better than humans,” said founder Joffe.
Boracas agrees that hiring with the help of AI is better than the status quo. The most responsible companies, however, admit they cannot completely eliminate bias and tackle it head-on. “We shouldn’t think of it as a silver bullet,” he cautioned.