Reducing Bias and Widening the Candidate Pool: Why We Built HireVue Assessments

May 24th, 2018
Loren Larsen, Chief Technology Officer
Artificial Intelligence,
Video Interviewing



In 2012, after almost a decade of doing on-demand video interviews for companies around the world, we set out to use machine learning to begin predicting from the interviews which candidates would be the best fit for a specific job. What resulted from that effort is HireVue Assessments, which collapse the traditionally separate interview and assessment steps into one step, improving the experience of candidates, and reducing both implicit and explicit bias in hiring.

Widening the Talent Pool with Video First

From our decade in working with recruiters and hiring managers at corporations around the world, we knew that HireVue video interviews were solving a big problem for both recruiters and candidates. For one thing, video interviews were freeing candidates to record the initial screening interview by smartphone or laptop computer at the times of their choice, without needing to travel to the recruiter’s office.

Additionally, HireVue video interviews were also relieving recruiters of the burden of ongoing scheduling demands and allowing them to review screening interviews at the times and places most convenient to them.

Even more importantly, our customers told us that on-demand videos were allowing their recruiters to consider more candidates than they were previously able to include and to offer each candidate the same fair, structured interview. This last point is key: being able to offer structured interviews that consistently deliver the same questions has helped our customers both widen the candidate pool and offer exactly the same engaging interview experience to all of them.

Solving the Dropout Problem

After helping solve the problems of scheduling screening interviews and helping recruiting teams consider larger talent pools, we kept hearing some consistent themes from customers. They still weren’t finding enough candidates to fill their roles, still learning that candidates did not like taking assessments or simply exited the process at that step of the process, and still taking far too long to fill their open requisitions, resulting in a loss of the best candidates and positions left unfilled.

These companies were not able to consistently hire the best people; as one of them told me, they were hiring the best people who were willing to endure their hiring processes.

Capturing the Screening Interview and Assessment in a Single Step

We recognized that there was a real opportunity to do even more than simplifying the interview process and making it more fair. Could we use our existing video interviews plus machine learning to provide a single experience for the candidate that would serve as both an interview and an assessment in one step? From the candidates’ perspective, they’d be taking an engaging video interview, but also being assessed in a robust and valid way in a single step.

This saves time for everyone involved. From the hiring company side, it dramatically shortens time-to-hire, engages candidates, and lets them retain the best talent in the pipeline. Because companies wouldn’t be losing so many people in their lengthy hiring process, their sourcing needs would also be reduced. From the candidate’s perspective, one brief video interview would cover both screening and assessments steps at the same time.

Reducing the Impact of Unconscious Human Bias

As we explored the potential to combine the interview and assessment into a single experience , we quickly realized that machine learning represented a very real opportunity to eliminate unconscious human bias in candidate screening, as well. So when we sat down to design HireVue Assessments, we wanted to do four things:

  • Ensure that there is a clear performance indicator that differentiates the strongest from the least promising performers.
  • Ask the right questions to elicit responses that can be measured and that are pertinent to predicting job performance based on IO psychology research
  • Have the machine notice everything present in the interview (what someone says and how they say it), and build a model that uses only the data points that help predict success in the job.
  • Actively and methodically audit the algorithms to ensure that they aren’t adversely impacting a group of candidates. Then we look for ways to remove features that may cause biased results.

These last two objectives were especially critical. Humans are biased – every one of us is. In the hiring process we all operate out of certain stereotypes and beliefs about what makes a good hire, but by training the algorithms on a job-relevant performance metric, we believed we could eliminate much of that bias in ways that human evaluation of interviews simply can’t. Today, nearly one million assessments later, our customers are reporting that this approach is working well for them, helping recruiting teams save time, hire the best talent much faster, and measurably increase the diversity of their hires.

In my next blog post, I describe our process for designing HireVue Assessment models that provide decision support to mitigate bias for recruiters at the first screening step.