Could Responsible AI Change Recruiting for the Better?

Artificial intelligence (AI) is influencing hiring decisions across the economy – regardless of whether we are ready for it. Adoption of AI recruiting tools has substantially increased in the past few years, coming in the shapes of resume parsers, predictive analytics, and more fringe applications such as performance simulations or inferences about the color of an applicant’s skin.

But the explosion in AI recruiting tools has not seemed to solve a great problem in the current hiring landscape – the challenge to hire “tech talent.” Last year, more than 250,000 data scientist positions went unfilled, and resignations in the tech sector remained higher than the national average. It seems strange that high-paying jobs, often in companies with flexible work environments, struggle to find and retain talent – especially now, when a significant portion of the workforce has sought to change careers.

Many have suggested that the difficulty in hiring high-skill tech workers can be explained not by an actual scarcity of those with the desired technical skills but rather by faulty recruiting practices – insufficient efforts to hire a diverse workforce, implicit biases in the interview process or a blinding obsession with “poaching” talent from competitors.

Given the growing reliance on AI recruiting tools, and the early signs that regulation is to come, now might be an excellent time to examine whether they too may be playing a role in the problem. AI tools could be using biased data in the underlying models, inappropriate or uninterpretable models that can’t be easily corrected, or worse – both.

However, there’s good news for AI enthusiasts in recruiting. As awareness of the shortcomings of traditional AI systems have grown, so has the innovation to create alternative forms of AI that are less prone to bias. One in particular stands out as a perfect fit for the recruiting space: the practice of Responsible AI (RAI). Yes, it’s a thing! And contrary to popular belief, it isn’t just regular AI with “ethics” in its branding – it is an intentional, growing practice with its own benefits and downsides, but one that we think is here to stay. So it’s worth our time to examine what exactly RAI is, why it can be so powerful in tech recruiting, and what kinds of questions one should ask vendors if curious or ready to incorporate it into the current tech stack. 

What is Responsible AI?

Responsible AI describes autonomous processes of decision-making with clear standards and protocols for ethics, efficacy and trustworthiness. In other words, it is a type of AI design that is serious and intentional about its underlying moral principles, such as transparency and impartiality.

Many cloud computing and AI providers have already begun offering solutions for RAI. Although it may have different names, in general, providers such as Amazon Web Services, Microsoft and Google are speaking of tools and methods to implement AI in a way that can be transparent, interpretable and scalable for different stakeholders.

There are downsides to committing to RAI. In general, you will not find it next to other trendy buzzwords such as “artificial neural networks” or “deep learning,” which can be attractive from a marketing perspective. There also may be no guarantee of the maximum possible accuracy rate because that is not the primary goal of RAI.

But the benefits of RAI in the recruiting space are too attractive to be overlooked. In fact, its focus on transparency and interpretability will likely lead to it becoming a required operation from auditors and regulation. In the context of today’s tech recruiting problem, RAI also stands out as an excellent complement to the rising practice of skills-based hiring, one of the most hyped ideas for bias reduction in tech hiring.

Why RAI is Perfect for Skills-based Hiring

There is a strong synergy between RAI and skills-based hiring – a method that de-emphasizes some aspects of an applicant, such as formal education, and focuses instead on objective and transferable skills. Organizations hiring this way are more likely to tap into “hidden talent” pools otherwise overlooked, such as refugees who may not possess validated educational credentials but who may nonetheless have the necessary skills for the jobs. Some studies suggest companies that are successful at tapping hidden talent could be 44% less likely to face a skill shortage in technical areas, which is why skills-based hiring has such a strong appeal to tech recruiters.

There are three ways in which RAI perfectly complements skills-based hiring: (1) It programmatically mirrors the practice of input restraint; (2) it provides recruiters with a technical basis for accountability; and (3) through its emphasis on algorithmic interpretability, it gives practitioners a better way to engage and troubleshoot their analyses. These are not easy concepts by any means, so it’s worth exploring them in a bit more detail.

Input Restraint

One major way in which RAI is compatible with skills-based hiring is its tendency to constrain itself when selecting appropriate inputs for the analysis. Rather than including every possible feature to maximize metrics such as the coefficient of determination, RAI actively scans for possible bias in the input space and removes bias-prone features, even at the cost of higher model accuracy. A classic example of this principle can be found in a hypothetical classification system for a loan – something that cannot by law be influenced by someone’s race, for example. But variables such as one’s ZIP code may be correlated with race, and this is something RAI would likely exclude from the model, whereas perhaps traditional AI systems may not. Race and gender bias in AI, by the way, is a well-documented phenomenon, due in great part to improper attention to this stage.

Decision Accountability

Another major point of compatibility between RAI and skills-based hiring is RAI’s facilitation of decision accountability. Think of it this way: If the ultimate hiring decision diverges from a suggested model that was purely based on skills, the decision-maker may owe the rest of the team a convincing explanation, whereas before, they may have gotten away dismissing the AI – particularly if the company is indeed employing skills-based hiring or other bias-minimization efforts.

Of course, there is a range of perfectly valid explanations for why one may not accept the model’s recommendations. The applicant may not have performed well in the interview, or maybe there were other barriers involved, such as immigration status or unwillingness to relocate. The point of implementing RAI is not to take away human agency. Rather, it is to incentivize the habit of transparently explaining a hiring decision in terms of its convergence or divergence from a model based on objective inputs.

Algorithmic Interpretability

Finally, RAI goes hand in hand with skills-based hiring because of its unrelenting commitment to algorithm interpretability. Many AI systems maximize high-accuracy predictions through techniques called black box models. These models have generated remarkable results in areas such as graphical processing and genomic analysis. But there’s a catch – many of these models are so complex that data scientists and business leaders cannot easily interpret them.

Interpretability is crucial for bias-free inference and arguably a moral need for problems whose answers may dictate someone’s credit score or job placement. RAI takes this principle seriously, often lending itself to a model choice that is inherently interpretable, such as a regression or decision tree. Moreover, many RAI solutions may come with data visualizations that can help stakeholders of different technical abilities understand and engage with the model decisions.

Is Your Tool of Choice RAI-compliant?

If you’re convinced that RAI is the way to go, you may be wondering whether your current tech stack is already equipped with RAI. It’s not so easy to know this right away because there are different standards for assessing RAI compliance. But there are great questions you can ask your vendors to gauge alignment to the principles of RAI.

• Ask about the data source. Was the data that was used in the analysis audited? Was it checked for explicit and hidden biases? Your vendors should be comfortable explaining their data provenance and should be able to demonstrate a legitimate effort in minimizing bias through one of many methods – input restraint (also called feature exclusion) being one of them!
• Ask about not only the “could” but also the “should.” AI can do so much. But being able to do something does not necessarily mean we should. Products that may attempt to assess compatibility through simulations or poorly understood validated features could put you at risk of passing on good candidates, and the possibility of facing a lawsuit or two. So ask not only how an inference is made but also why it should actually help you in the first place. Being accurate is not good enough.
• Ask about the model choice. Is your provider’s website filled with AI buzzwords? Is the only reported metric a single accuracy value? If so, challenge your provider! Ask how the “extra accuracy” in the models will help with your hiring needs. And don’t forget to ask how your provider can help you interpret the model results; after all, checking for bias should not be a one-time exercise. 

Conclusion

At a time when higher salaries and flexible work environments are of high importance to job applicants, it shouldn’t be this difficult to hire and retain technical talent. But a combination of antiquated hiring practices and unhelpful AI practices has led us here. Thankfully, the practice of Responsible AI is growing, and its adoption may change the tide for tech recruiters. RAI’s intentional focus on transparency and its compatibility with skills-based hiring may greatly reduce bias in tech hiring and can give us long-lasting impacts for a future in which career transitions to the tech sector become even more common.

Next time you consider an AI recruiting tool, ask not only about its accuracy but also about the underlying methodology and value to a fair and transparent hiring process.