Proxies in eLearning Data Reveal Promise, Pitfalls of AI

Working with data to better understand learners, determine(or demonstrate) the effectiveness of training and performancesupport tools, or figure out which applicant to interview, hire, or promote canbe efficient and effective—if you have good data. Unfortunately, much of whatis regarded as “big data”—or any size data—consists of proxies. The problemwith proxies in eLearning data is that proxies often miss the mark, omittingcrucial information, or focusing attention on the wrong details.

Codified opinions

Proxies are models. They are stand-ins for facts. A proxymight consist of data about actual people who are similar to the personwhose learning needs you’re trying to anticipate. Or, it could be informationthat is easily available or legal to gather, like a person’s ZIP code, thatproxies information—intentionally or not—that cannot easily or ethically beconsidered, like race. Most insidiously, proxies can include patterns that an AI(artificial intelligence) algorithm has detected and is using,without the knowledge or intention of the humans using the AI-poweredapplication. AI refers to computers performing in a way that “mimics someoperations of the human mind, such as making decisions based on data,recognizing objects and speech, and translating languages,” according to JaneBozarth, the Guild’s research director and author of ArtificialIntelligence Across Industries: Where Does L&D Fit?

Proxies are ubiquitous because people building data-based,automated tools “routinely lack data for the behaviors they’re most interestedin,” Cathy O’Neil wrote in Weapons of Math Destruction. “So theysubstitute stand-in data, or proxies. They draw statistical correlations betweena person’s ZIP code or language patterns and her potential to pay back a loanor handle a job.” An obvious—but increasingly common—problem is that many ofthe correlations lead to discrimination.

A solid data model, which should be the foundation ofany AI-basedeLearning or performance support tool, requires a large amount ofgood, reliable data. The developers behind the model, and the L&D teams andmanagers using its outputs, need a steady stream of real data. Adding new dataallows them to constantly test the model, get feedback on how it’s working, analyzethat feedback, and refine the model. Off-the-shelf, “black-box”AI algorithms lack these features.

“Models are, by their very nature, simplifications,” O’Neilwrote. “Models are opinions embedded in mathematics.” The programmer whocreates the model decides what to include and what to leave out. Sometimes theomissions are not harmful and are even necessary and helpful: O’Neil providesthe example of Google Maps, which “models the world as a series of roads,tunnels, and bridges.” The details—landmarks, buildings, traffic lights—thatdon’t appear don’t make the model any less useful for its intended purpose.

Not all omissions are benign, though; many models sacrificeconsiderable accuracy in their quest for efficiency. An algorithm that filtersapplicants for a job might improve efficiency by quickly narrowing theapplicant pool to a manageable number of individuals to interview, but if itfilters out all individuals who live outside a specific area or whose resumesdon’t specifically mention a software package, the best candidate might bepassed over.

In some cases, the difference between whether a model ishelpful or harmful depends on what it’s used for.

Consider how proxy data will be used

Many large corporations use automated testing to screen jobapplicants. Some of these tests, which resemble well-known “five-factor”personality tests, purport to be able to predict which applicants will performbetter on the job or are less likely to “churn”—that is, which applicants mightremain with the company longer. Obviously, actual data about the individualapplicants’ performance on this job, for which they haven’t yet been hired, isnot available. Neither is actual data on these individuals’ historical workperformance. So, the automated hiring and onboarding programs use proxies topredict the behavior of applicants and rate them, relative to others in theapplicant pool.

While there is a role for predictiveanalytics in eLearning and in personnel decision-making, thosepredictive analytics should use solid data. Instead, they might rely on theanswers to unscientific questions that often force candidates to choose betweentwo illogical and unsuitable responses. One example O’Neil provides is choosingwhich of these statements “best described” them:

• “It is difficult to be cheerful when there aremany problems to take care of.”
• “Sometimes, I need a push to get started on mywork.”

What if neither describes the applicant? Or both? Thealgorithm-based tests are rigid and lack any appeal or “sanity check,” such asa human with whom to discuss responses and results.

A critical problem with automated screen-out algorithms isthat there’s no way to tell how many rejected candidates would have beenstellar employees, information that would provide valuable feedback forimproving the tests. This lack of feedback is problematic for both theapplicant and the hiring organization: Rejected candidates have no idea whythey were rejected and cannot appeal. The hiring team has no insight into whythe candidates who “passed” were selected.

But the tests are not inherently problematic; it’s the useof them to screen job applicants that is frustrating for applicants, potentiallydiscriminatory—and extremely ineffective, according to the HarvardBusiness Review. But if the tests were used with existing employees,they could enhance team building and communication. “They create a situation inwhich people think explicitly about how to work together. That intention alonemight create a better work environment,” O’Neil wrote.

She suggests using automated testing to gain insights, notto screen out applicants. Rather than analyzing the skills that each applicantbrings, O’Neil wrote, these tests judge applicants by comparing them with“people who seem similar.” They tend to identify people who “are likely to facegreat challenges,” she wrote—in school, in a job—and reject them.

Instead, the information gleaned from applications could beused to identify the services and programs applicants need to meet theirchallenges and succeed. If managers used automated questionnaires to learnabout new hires during onboarding, say, rather than to filter out nonconformingapplicants, the actual data they’d have about people they’d hired could helpthem improve workplace practices. Managers could use this data to tailortraining, create performance support tools, and shift workplace culture and policiesto ensure the success of a broader pool of applicants and new hires.

No proxy for soft skills

Finally, an additional problem O’Neil identified with usingautomated testing and proxy models to screen employees and applicants is theirinability to capture essential skills. Employees have value far beyond theiryears of schooling or the number of phone calls they complete in an hour, but themodels have no digital proxies for soft skills. The inherent value in being theteam member who coaches the others or defuses tension in group meetings isimmense—but impossible to quantify. Teams need all kinds of people, and not allskills are countable or digitally measurable.

As with predictive analytics, AI-powered screening tests are a usefultool, but only when used as part of a system that includes feedback, analysis,and collaborativedecision-making by managers who know the individuals they areevaluating. When used for suitable purposes and with appropriate humanintervention, proxies in eLearning data collection and analysis can enhanceefficiency and effectiveness.