Understanding AI Bias at HiringBranch

Having recently passed the New York City Local Law 144 AI audit, I sat down with Chief Research and Development Officer at HiringBranch, Assaf Bar-Moshe, PhD, to talk about bias in the hiring process and in technologies like hiring assessments. From human judgement in hiring to a new generation of AI assessments, Assaf is a true authority on the topic. Check out the full interview below.

Chandal: “Assaf, you told me that you believe AI can reduce the bias of individuals rather than introduce it. Tell me more about that.”

Assaf: “Bias is a big issue not only in the AI world but also when assessments are evaluated by humans. From this point of view, AI actually reduces bias. AI is blind to accents, the color of the person, the gender of the person, their age, and other features. As data scientists, we do not introduce these features into the machine, so the machine does not use them for classification, and therefore, we don’t look at a group of women and men separately, for example. We don't measure native speakers and non-native speakers separately. The machine makes decisions without knowing if somebody’s a man, is white, or anything like that. 

"As much as they would like to, humans cannot entirely isolate their prejudices. This is more pronounced if the team has a large volume of assessments..."

Chandal: “So you think when a test is evaluated by a human, they introduce bias into their evaluation by nature?” 

Assaf: “As much as they would like to, humans cannot entirely isolate their prejudices. This is more pronounced if the team has a large volume of assessments, which would not be evaluated by a single person. Having different people performing evaluations means different opinions and different skills will be introduced into the process. No matter how you try to standardize the “rubric” or the evaluation process, there will be bias that gets into the evaluation process when it’s done by humans. Humans can have different opinions about the world around them, and they will possess different soft skills. They will also have a bias towards different accents or a different tolerance towards grammatical mistakes, and so on.”

Chandal: “Do you think AI can be used to enhance what is being evaluated by assessments?”

Assaf: “AI enables an evaluation of dozens or hundreds of different features. So, if we are evaluating communication skills like grammar or fluency, there are dozens or hundreds of features analyzed at the same time, something that is very difficult for a human to do. Humans can say the same thing in many different ways. Some of them are unpredictable. As such, if they were not modeled before, they wouldn't be evaluated. Using machine learning, however, the models are more flexible and allow us to account for things even if they were not modeled in so many words before.”

"A multiple-choice format is not AI. This is just straightforward true or false. So the only way that an assessment can truly be AI-driven is if you have open-ended questions."

Chandal: “How do you think AI is shaping a new generation of hiring assessments?”

Assaf: “So with the power of AI, we always learn, and we can build dynamic evaluation processes. Rather than a standardized rubric by which our assessments are evaluated, AI allows teams to have rubrics that get refreshed and updated on the go, all the time. If new material comes in, the machine learns from this new material and becomes smarter. Rather than being static and saying “this grammar mistake is very bad”, it might be the case that with time, the machine learns that it is not that severe. So we have the ability with AI to maintain a dynamic system of evaluation, rather than a static one that can’t evolve with time.

Chandal: “For candidates, do you think AI enhances the format of the evaluation?

Assaf: “Many communication assessments are closed-ended, which flattens the candidate's abilities and opens the door for cheating. Using a dynamic AI system allows hiring teams to construct assessments with open-ended questions, where the candidate needs to react to a given scenario in their own words in writing or speaking.”

Chandal: “So what is your opinion on multiple-choice 'AI' assessments?” 

Assaf: “A multiple-choice format is not AI. This is just straightforward true or false. So the only way that an assessment can truly be AI-driven is if you have open-ended questions. 

Instead of measuring one right or wrong response, AI measures the way things are said, how they are said, grammar, fluency, how relevant the answer is to the question that was asked, whether the candidate showed empathy to the customer, used proper language, and more. It’s even better if the open-ended questions mimic real on-the-job scenarios. Regardless, with open-ended questions, there is no right answer, and therefore, it’s nearly impossible to cheat. 

"With enough data and confidence in the data, management teams can use it to make decisions about the business, like which location candidates perform better or which hiring partner brings more revenue"

Chandal: “With AI, do you think hiring teams can get meaningful performance data?

Assaf: “With AI, there is no limit to the number of assessments a team can evaluate at the same time. This scales up and yields a lot of data over time, which can be used to improve the reliability and the accuracy of the assessment, eliminating false positives with time. 

With several years of experience, we were able to get to a point in which we basically have almost zero false positives in the HiringBranch assessment, meaning no unqualified people passed the test. We have very few exceptions, if any. So hiring teams can actually become more confident about their performance knowing that people who passed the test do a good job. 

With enough data and confidence in the data, management teams can use it to make decisions about the business, like which location candidates perform better or which hiring partner brings more revenue, etc. Using AI to process large amounts of data also means decision-making can happen faster. 

Chandal: “Any last thoughts on adopting AI for unbiased hiring assessments?”

Assaf: “As a scientist, I’m excited by the potential of AI that we’re already seeing within the recruiting and talent acquisition space. It’s certainly increased the hiring confidence of our customers. If you’re thinking about adopting AI in your hiring pipeline, just be cautious, do your research, and verify that what the service provider sells works, as proven by data and analysis. And when it does, just embrace it. It will save so much time and effort while giving you unbiased results and more qualified candidates.”

‍

‍

Image Credits

Unsplash/Brian Suman