Recently it was reported that Amazon had scrapped their machine learning recruitment system because it was favoring male candidates. Their machine learning team spent years developing a way to automate the hiring process using a decade worth of past Amazon employee resumes; they would soon learn in doing so the system had taught itself to strategically weed out female candidates. For instance, any resume with the word ‘women’s’ in it was immediately downgraded and scored lower.
What gives? Learned historical data based on past hiring decisions will always produce a biased system because human beings are inherently, and (mostly) unconsciously biased.
Using our own AI technology, TalentWorks pinpoints the biases of hiring managers by analyzing and sampling 100,000+ jobs from our index of 91 million job postings. In doing so we’re able to identify the norms and outliers of the industry such as the number of applications per interview and how that relates to the greater labor market. Example: Racial bias.
How much does race still matter in the US job market?
After analyzing thousands of job applications, outcomes and applicants, we discovered several key things:
- Non-white job applicants got 2.3x fewer interviews than their white counterparts;
- For non-white job applicants, if a resume mistake reinforced a racist stereotype, it basically disqualified them.
Through our data, we’ve found the following contributes significantly to combatting racial bias:
- Limit the number of collaboration-oriented words in your resume, such as, “team player”, “helped”, and “assisted”. Doing so will improve your chances for objectivity by 63%.
- Anchor your experience by using industry buzzwords and acronyms. This increases objectivity by 34%.
- Use concrete numbers; specifically, for every 3 sentences use 1 number to demonstrate your impact. Especially for people of color, quantifying the impact that you made with numbers helps remove subjective bias (+23% boost).
Returning to gender bias, we’ve actually found that resumes with obviously female names had a +48.3% higher chance of getting an interview. Names such as ‘Monica’, ‘Zoe’, ‘Ashley’ and ‘Evelyn’ had a significant boost over men comparatively.
This isn’t incredibly surprising when compared to what happened at Amazon, another tech company with a stark talent problem (men make up 73% of professional employees and 78% of senior executives and managers). The greater labor market suggests that there is an immense benefit for hiring women. Women are outperforming men in school, and most recruiters are women (who want to support other women).
‘Tech’ is just one industry Talentworks analyzes. Our data is based on thousands of applications, applicants and outcomes across 681 cities and 140+ different roles/industries. Artificial intelligence and deep learning are the future of recruiting. We hope to empower jobseekers to find their ideal position.
First, we randomly sampled 100,000 jobs from our index of 91 million job postings. We extracted the number of years of experience, job level and employment type for each job using TalentWorks’ proprietary parsing algorithms. We then used a blended Gaussian-linear kernel to calculate experience densities. Finally, we used an averaged ensemble of multiple independent RANSAC iterations to robustly calculate inflations against outliers. This was done in python with pandas, sklearn and scipy and plotted with bokeh.
Why Are We Doing This?
We can boost the average job-seeker’s hireability by 5.8x right now with ApplicationAssistant. But, what makes ApplicationAssistant work has been an internal company secret until now. We’re fundamentally a mission-driven company and we believe we can help more people by sharing our findings. So that’s exactly what we’re going to do.
We’re not only sharing this data but also sharing all of it under the Creative Commons Attribution-NonCommercial-ShareAlike license. In other words, as long as you follow a few license terms, this means you can:
- Share: Copy, redistribute the material in any medium or format.
- Adapt: Remix, transform, and build upon the material.