Machine learning (ML) is currently a hot topic across many fields, from medicine to IT to engineering. It is also increasingly being used in HR, particularly in hiring processes – from resume scanning to assessments to video interviewing. New advances suggest that ML can dramatically improve rates of making successful, diverse, and quick hires. Removing bias in the hiring process is one of the most exciting potential uses for machine learning. What does machine learning mean exactly, though, and can it really ensure an unbiased hiring process?
Simply, it is a type of Artificial Intelligence (AI). AI is a branch of computer science that gets computers to perform tasks that would normally require human intelligence. ML is one of several ways to do AI, and it involves writing a computer program (an ML model) that will recognize patterns in data that might be useful and find any correlations for you. What then? If an ML model can be programmed (or trained) to reliably recognize patterns in data, then the same ML model can be used later to predict information given some new data. In the example below, let’s imagine a company wants to predict the intelligence of their job candidates based solely on their favorite ice cream flavor and their years of work experience.
How does Machine Learning Work?
Most ML models used today are based on Supervised Learning. In these cases, a human guides the ML model by providing data in the form of input-output pairs during the training. Using the ice cream/intelligence/work experience example, the inputs would be a candidate’s favorite ice cream flavor and their years of work experience, and the output is how the candidate performed on an intelligence test. In supervised learning, the ML model gets to see inputs and output data while getting trained.
ML models also need a representation for the data that allows the computer to learn the correlations between the inputs with the output. Usually, the biggest challenge in ML is finding a good representation for the data. The representation could be very simple or complex, but to be useful, it needs to sensibly capture the essence of the correlation. For example, a useless representation would be the number of letters that spell the ice cream flavor (for example: vanilla = 7, chocolate = 9, strawberry = 10).
Once an ML model is trained, the model can only be trusted to make interpolations in the data and not extrapolations. This means that a good ML model that finds a reliable correlation based on training data with mostly vanilla, chocolate, or strawberry ice cream lovers should work for other similar candidates who are added to the data set after the training. However, it won’t necessarily work for candidates who like butter pecan ice cream and it definitely won’t work for people who say they like pizza.
Why is all this important? Because it shows that the success of a machine learning model is completely dependent on the data set used, as well as on the guidance it receives from humans during training. In models based on Unsupervised Learning, some of this bias can get removed by having the ML model identify patterns between only inputs and not outputs, but predictions and understanding are harder to determine.
You might also like: What is Natural Language Processing and Should You Be Using It?
So, Can Machine Learning Remove Bias in the Hiring Process?
The short answer is “not yet” – unless you want a computer program to make completely random hiring decisions for you. Until humans are no longer involved in defining what successful job performance is, there will be bias in the hiring process. Let’s go through an example.
A job description is the hiring starting point. Who writes it? How is it written? Was it written to reach a diverse audience? Why are certain skills required while others are nice to have? Even if you’re using an ML tool to match resumes to your job description, you’ve already introduced some bias by having humans involved in creating the job description.
A behavioral assessment is used early in the process, which is a great idea! Assessments are useful tools to objectively measure attributes which are important in a role. But to decide which assessment to use and how it will be scored, you again must get humans involved. You can use ML to match up assessment results with the target job, but there may already be bias in having chosen what competencies are important in a role.
A video interview is the next step in the process, and an ML program can score a candidate on different characteristics based on their responses, which is pretty cool. In order to learn how to do this, though, the machine learning needs human inputs. How else will the computer know whether a candidate is giving a favorable answer to a question?
And Now the Good News
ML is not an elixir to cure all that’s wrong with hiring processes, but it can help us make hiring decisions with less bias as long as the inputs used have as little bias as possible, too. What else can help? Doing a thorough job analysis before putting any kind of hiring process in place, using objective hiring steps (like assessments), and making sure that your human interviewers are well trained and well prepared. Although these aren’t shiny new things, they can still be impactful in any hiring market.