Minneapolis Public Schools is up against a vicious cycle of sorts.
Like a lot of districts that are relatively poor (65 percent of students qualify for subsidized lunch) and nonwhite (70 percent of students are people of color), it has a hard time hanging onto teachers. Its turnover rate in 2018 was 10 percent, as compared to the national average of 8 percent. It was even worse for teachers of color: 12 percent.
The constant drain means Minneapolis’ teachers are, on average, less experienced and less diverse. It also usually ends up costing the district a lot of money. More importantly, it leads to poorer outcomes for students.
“There’s a lot of research that shows there’s a lot of variation in the effectiveness of teachers,” labor economist and University of Minnesota associate professor Aaron Sojourner says. One year with an inexperienced or less skilled instructor can make a noticeable difference in a student’s life – including her future education level and earnings.
Sojourner and his fellow researchers got to wondering if some of this turnover could be avoided by hiring smarter. Maybe, with the help of algorithms and AI, they could find the patterns in resumes that lead to longer-lasting, better teachers, and avoid more human errors, assumptions, and biases in the hiring process.
So they gave it a shot. They took more than 16,000 applications received by Minneapolis Public Schools between 2007 and 2013, and used a machine learning program to analyze patterns in work history, experience, and how well they fared in the district. The results were published last month in Applied Psychology.
Part of what they found wasn’t that surprising. Teachers with more work experience tended to be better hires and stay on longer. What was surprising was the range of what that work experience could be. Sima Sajjadiani, University of Minnesota Ph.D. and lead author on the study, says job titles mattered a lot less than the knowledge and skills accumulated in previous positions.
“Waiters and waitresses, people with psychology backgrounds, people with sales experience,” she says—these backgrounds were all “actually helpful” in creating better teachers, even if they’ve never been in charge of a classroom before.
“There wasn’t any extra bump for having experience literally as a teacher to jobs that are very similar to teaching,” Sojourner says.
Another surprising element: attitude. Candidates who used phrases like “passion” in their applications tended to stick around longer and do better than those who didn’t. These are terms that can sound like clichés to hiring officers, who might even choose to ignore them entirely—but by the numbers, they may matter.
This isn’t the first time machine learning has been used to optimize the hiring process, and it hasn’t always gone well. The most famous example was a computer program used by Amazon until 2015— when it was discovered it had developed an accidental gender bias by analyzing way more applications from men than from women.
There are a lot of concerns about allowing computers to run away with things and introducing dangerous (if unintended) consequences. Sajjadiani thinks they're justified.
“People ask me if I think hiring should move completely to AI and machine learning, and I say ‘absolutely not,’” she says. There always has to be a “human” voice in the room to make sure that data comes with the context necessary to see the whole picture.
But the results of this study show promise. In fact, simulations suggest that the recommendations from the algorithm would decrease racial disparities in Minneapolis Public Schools’ hiring process. Sojourner says the next step is to try the same study elsewhere, get more data, and see if they get a similar outcome.
Human judgment will always be required. But maybe we’ll judge just a little bit better with a robotic angel on our shoulders.