Don’t Rely on Data Science and Machine Learning To Hire (Yet)
AI systems are all the rage but they have serious inherent flaws
Pressed for time and budgets getting tighter, companies have to devise or utilize innovative ways to win the war for talent. McKinsey predicted that the world will have a shortage of talent — 40 million too few college-educated workers, to be exact. ManpowerGroup’s recent Talent Shortage Survey found that 38% of employers have difficulty filling jobs. With the job markets getting more competitive, companies are turning to the latest developments in technology: artificial intelligence systems to source, screen, and hire the best candidates.
It’s a small wonder why: hiring in 2020 and the coming years introduced a paradigm shift for companies. Employees are no longer as loyal as before. Millennials and Gen Zers, who have such traits, are gradually occupying more positions in the workforce. To reverse this trend, companies are focusing on designing a compelling employee experience.
Yet, incumbents are often resistant to change. Most of the time, they are too heavy to move fast, since their conventional ways of working in silos still exist. To change the way gears are moving, companies are increasingly turning to HR-tech vendors, driving the HR-tech market to be worth at least USD$3.6 billion by 2024.
This resulted in a groundbreaking development: companies are increasingly using artificial intelligence and data science to determine their hiring decisions.
It initially made sense. These algorithms, built by data scientists and hiring managers, are meant to eliminate human bias. Traditionally, humans have biases and predilections, which makes our hiring decisions unfair — it is the reason why you would fare better in an interview if you have the same hobby with the interviewer. It builds rapport, which can cloud their evaluation after the interview.
Algorithms remove that clouded judgment. After all, computational algorithms often outperform human judgment, which means we often trust it more than ourselves.
Using data, the algorithm assesses objectively without the cognitive biases we all have. What could be a two-hour job of screening candidates can be reduced to a few seconds: to a company, time not wasted is money saved.
Unfortunately, saving the few hours might result in more money lost as critics are showing that algorithms are in fact, increasingly biased.
In 2003, a team of researchers at the National Bureau of Economic Research conducted a field experiment. They sent fictitious resumes to job ads in Boston and Chicago newspapers, some assigned to an African-American-sounding name and others a White-sounding name.
In the end, recruiters overwhelmingly favored the applicants with a White-sounding name.
In part, algorithms are our opinions embedded in code. When there are human bias and prejudice, the algorithms become biased, leading to machine learning mistakes and misinterpretations. If you were to create an algorithm around the recruiters’ decisions in the experiment above, the algorithm will learn that a “good” hire has a White-sounding name — an algorithm is only as good as its data, and biased data leads to biased hiring decisions.
Amazon was the victim and culprit of that.
Amazon relied on an internal ‘secret’ hiring algorithm, that was created in 2014, to assess candidates based on their traits relative to the top performers in the company. It was scrapped three years later when they discovered that it overwhelmingly favored males: the company’s top performers were disproportionately males (which could be a result of inherent gender bias).
This made the recruitment world wonder: are hiring algorithms truly worth it? It can be a yes, but only if companies know its inherent flaws and how they mitigate it.
Never Fully Trust Algorithmic Recommendations
Companies often fall into the trap of relying on algorithms to make hard screening decisions. Yet, it’s clear that algorithms are partly human decisions in code, which means that there’s going to be human bias along the way. When companies like Amazon created such an algorithm, they are infringing on legalities and moralities by being biased against women.
Essentially, algorithms are not neutral. As a recent study by Cornell University shows, hiring algorithms are too opaque to tell if they are fair or not.
What do vendors mean when they say fair and unbiased? What are the criteria involved to determine whether they are judging candidates ethically? Are they putting in effort to ensure that it is truly objective, rather than just leave it be?
As such, hiring algorithms are still not the panacea that companies are hoping for in the war for talent. Companies need to be wary of trusting these algorithms too much: if we rely on them and make biased decisions all the time, we’ll only make more biased decisions in the future. This time, they’ll only get faster.
Humans Must Always be Behind the Wheel
Tesla’s autopilot feature meant that cars can always automatically steer, accelerate, and brake automatically within its lane. It can also navigate and change lanes. Recent developments have allowed it to navigate complex, tighter roads as well.
However, it’s not full autonomy.
Tesla clearly states: “current Autopilot features require active driver supervision and do not make the vehicle autonomous”.
Just like Tesla’s autopilot feature, hiring algorithms are the same. We are decades away from reaching fully autonomous vehicles and that goes the same for hiring algorithms too. Good algorithms exist but the success doesn’t happen by accident. Rather than over-rely on the algorithm early on, companies should audit and modify them to prevent inequities from perpetuating.
Hence, hiring recommendations should be rigorously reviewed. Hiring managers must treat such algorithms as an aid, not a cure-all.
Audit its Fairness with Human Review
For companies, the benefits of speeding up time-to-hire and streamlining processes can often outweigh the risk of inadvertent discrimination.
To counter that and anticipate potential algorithmic discrimination, companies can run human reviews. For instance, they can conduct manual reviews of the correlations that the machine learns and selectively eliminate any that may appear to be biased.
Companies can also periodically carry out random spot-checks on machine resume decisions: put algorithms through an extensive human review to see who the algorithms select and why. The aim is to uncover potential causes of bias.
Audit its Fairness with AI
It sounds ironic, but algorithms can also help to identify biases in an algorithm.
At Penn State University and Columbia University, researches developed an AI tool to detect discrimination based on attributes. HR tech startup Pymetrics open-sourced Audit AI, an algorithm that also determines whether there is a systematic bias in an algorithm (although it covers a broader spectrum of algorithm types).
Companies can also use multiple algorithms to help limit blind spots. This way, companies can exclude a qualified candidate simply based on a single metric (which algorithms often use, since it ends up with a single score).
Data Needs to Have Context
Not every algorithm is created equal. To create a good algorithm, data scientists must understand the variables in the context of employment.
For instance, commute distance is often cited as a factor in determining work performance. Data shows it can negatively affect work performance and increase absenteeism. Yet, commute distance is governed by our place of residence — which is governed by housing prices, income levels, and even race. An algorithm taking the commute distance into account at the surface level may be biased against protected groups like racial minorities.
Without context, data will not be enough to predict a person’s potential at work. To circumvent this, hiring algorithms often scrape for more data. Yet, this might not work as intended: past performance may be potentially firm-specific, which defeats the purpose of getting more data.
Fairness can Become A Struggle
A VU University Amsterdam study published December last year showed that while AI may not necessarily improve nor degrade ethical values, it can shape what we believe to be ethical in the first place.
In the study, an AI recruitment tool used by a multinational corporation was studied. The tool replaced standardized online tests with neuroscience games and automated video analysis. Over 7 months, the tool was piloted and implemented, aiming to objectively measure over 120 skills and traits of candidates. It then recommends candidates to hiring managers, who will bring them for an in-person and subsequently, a panel interview.
Eventually, candidates who fared worse than expected tried to game the system by cheating on the neuroscience games.
Team managers were also skeptical of the hiring algorithms, citing that they might be hiring more homogenous employees. Some of the hires they wanted also failed the AI assessment, yet nail the human assessment.
The AI team also met with pushback on whether the algorithm works or not. For instance, those with high scores on the AI assessment failed the subsequent human assessments. HR managers argued that the AI assessment might be failing — the AI team then argued that the human assessors were biased and thus, incorrect.
The power of algorithms and making data-driven decisions is something companies must harness today. Like McKinsey’s prediction, the war for talent has occurred — exacerbated by a younger workforce and paradigm shifts in mindsets. To screen faster, hire faster, and still increase the quality of candidates, hiring managers will wind up turning to HR tech vendors or developing algorithms in-house (if they have the time and resources).
Hence, companies need to consider a few factors when using an algorithm to make hiring decisions:
- Have consensus company-wide as to what is considered fair. As the VU University Amsterdam study showed, candidates, hiring managers, team leaders, and the AI team all share different views on what is considered fair. No one wants to admit that they are wrong, and they are quick to point fingers as to who is at fault for the poor decision-making.
- Use algorithms as a guide and not as a pill. Algorithms are not inherently neutral. Having our human opinions encoded means that it’s essentially a digital human with a single line of thinking: how much does this candidate match with my criteria? If we over-rely on the algorithm, we run the risk of making wrong decisions.
- Algorithms can lead to homogeneity. Suppose your algorithm determines that candidates with attributes A and B are great hires. You decide to hire such candidates, which feedbacks to the algorithm and reinforces their choice framework. Over time, you’re going to hire candidates that are all holding attributes A and B, which means the company is going to be lacking in diversity — not just of race and gender, but also of thought and leadership.
- Invest in auditing teams and algorithms. Use a multi-pronged strategy. Systematically improve the algorithm by using technology and human reviews. More time invested can potentially improve the quality of hires in the long run.
- Trust your judgment and question everything. Don’t listen to HR-tech vendors too whole-heartedly. Though they have a ‘proven’ algorithm, you still have to think in the context of your team and company. Question the algorithm’s training corpus and ask the right questions: are they actively combating bias in their algorithm, or are they simply piling data on data?
Will hiring algorithms eventually be trained to root out human bias? The answer is yes, but it’s too soon to tell. There might be a day where hiring tools will fairly and equitably hire people, but unfortunately, the devil’s in the details. Perhaps in time, we might see a day where a company focuses on the details, fixing bias in hiring with machine learning like every HR professional yearns to do.
Want more of such articles? I’m the founder of Human+Business, a leadership and human-focused publication on Medium. Read more here to learn how we can change the way we manage employees the way that it should have been.