Why Artificial Intelligence and Machine Learning May Not Give You What You Think They Give You

“Yes I love technology. But not as much as you you see. But I still love technology always and forever.”

-Kip Dynamite

Artificial intelligence (AI) and machine learning (ML) have been getting a lot of attention in the world of business and organizational analysis. Rightfully so, as there is potential to automate and speed up processes. PWC predicts that by 2030, AI will provide $15.7 dollars in global economic growth. But as with all tools, inappropriate use of these tools can actually be harmful. So, what do you actually get with an AI or ML system for something like employee selection? How do you know if it is the right fit?

The Society for Industrial-Organizational Psychology (SIOP) recently published a new white paper called “Algorithmic Justice” by 6 authors from Development Dimensions International. Sorry to not list them all but I’m trying to make these shorter. I wanted to highlight it in this post and try to talk about the concepts it covers in laypeople terms, because it brings up some very important points and I feel that sometimes are lost in these discussions. That white paper and others (including one other on AI in talent assessment) can be found at this page: Resources We Love.

Big shout out to the people DDI who put this white paper together! And no, I am not getting paid or sponsored by DDI to say that. Give credit where credit is due though…

I have heard several questions about these analytic methods and at times I think that people have felt that I am resistant to these newer methods. I probably have been a little defensive at times in large part because people tend to adopt the fancy new thing without thinking about what it actually does. I want the things I recommend to people to be a legitimately valid and applicable tool. And some of the promises that have been made for AI and ML are simply impossible to obtain. That doesn’t mean that these can’t be incredibly useful tools though. My hope is that by having more open discussion about things like this, we can be more educated on these tools and eventually satisfy all the needed demands. I’m cautiously optimistic that these are can of huge value.

Here is the big idea: We have to apply the same standards of fairness and accuracy we place on other workplace decisions to the use of these algorithms. I will highlight portions of the white paper and provide my own thoughts.

Bias in Artificial Intelligence and Machine Learning Algorithms?

The white paper begins by outlining AI and ML. The big advantage they mention is “…faster and accurate prediction of human characteristics…” (pg. 2) than humans can do. The way I have heard this described by laypeople is that it eliminates bias and allows us to remove humans completely from any business evaluations. I have two immediate thoughts on these claims. First, bias cannot be eliminated. Second, you don’t want to remove humans from the process completely. You do want to keep biased decision-making to a minimum, which is possible without AI and ML.

First, bias comes from incorrect human assumptions. Those human assumptions are inherent in data since we as humans make decisions about which data to include and how we use them. So although the algorithm’s decision-making process cannot be biased since it follows a program with rules to analyze the data, the human(s) who created those decisions rules still impose bias on those rules that the algorithm uses. We have to remember that anywhere humans are involved, they come with a history that colors their perspective, and as one of my former BYU professors put it, “When was the last time you ever did anything outside of your own experience” (he may have been quoting someone else). Any bias that humans have can be passed on to algorithms. The authors correctly caution on this point and note some of the good and bad ways that machine learning specifically has been used with corresponding consequences (see the table on pg. 3).

The second immediate thought I had is that you don’t want to eliminate humans out of the process. For one, requirements change and the algorithm will need to change to meet those needs. Another reason is that algorithms need extensive testing to see if they meet the intended use. One point the authors made was that assessment results should still be reported by a person to “allow a genuine human-human interaction”, which may be important for applicants. And lastly, not every single human experience can be measured well. Humans have to decide the threshold of acceptable error and appropriate use. Without that human piece to it, the algorithm becomes a bit aimless, and that’s not what you want in your business processes. Simple things like rater training, for example, can reduce biased human decision-making.

Using Artificial Intelligence and Machine Learning Well

I am glad that the authors take on this topic in the back half of the white paper. Because even if the promised results are a bit exaggerated, there could still be value in a new form of analysis. The authors of this white paper make an important point: any new method of analysis or evaluation still needs to meet the standards of responsible use and accuracy. Specifically, AI and ML applications need to have evidence of two things: 1) no adverse impact and 2) it can predict performance. Knowing these two things means that we need to know all the variables (or features) that go into an AI/ML algorithm, the way that those variables were weighted and analyzed, and how the algorithm is being monitored once implemented. This advice comes with good reason.

I have had conversations with some data scientists who use different types of machine learning methods in their work. They have expressed the idea that if demographic variables like race and gender are large contributors to a model for selection, then we should absolutely be using that information because it makes the algorithm better. In these same conversations with data scientists, I have heard them say that an algorithm is functioning really well and has very little error so it is ready for use.

It’s not that those ideas are incorrect. But the context in which we are working demands that we have more than an algorithm that just works in it’s own bubble. Sometimes, the algorithm itself becomes the priority instead of the decisions that needs to be made with the algorithm. By law, we cannot use information on those protected classes as any part of workplace decisions. This includes variables that may disproportionately pertain to certain races or gender (e.g., zip codes tend to have large percentages of specific racial groups in them). We need to make sure that the math works out along with the prediction of performance without including those protected classes. Just because the algorithm works doesn’t mean that the variables predict job performance. Require the algorithm creators to provide that connection to job performance and to explain how they arrived at that connection at each stage of their testing.

There is a difference of perspective and approach on these issues. Data scientists often are engrossed in the actual math of the algorithm. Consultants like myself are much more concerned with the process of arriving at evidence-based decisions. The two can work together, and when they do it can be extremely effective. I know that is overgeneralizing a bit, since people with my same PhD training are in data scientist roles and some teams of consultants include data scientists. But the slight difference in approach can change the ultimate outcome and focus of the work.

One Other Consideration

The authors give an interesting point toward the end of their white paper about using the AI and ML, particularly in selection contexts. They note the implementation of consent to have their information evaluated by an AI algorithm. Illinois apparently made it a law that applicants need to give consent to having their information analyzed with AI. I don’t fully know exactly where I stand on that, but it’s an interesting consideration at this point given some of the fine tuning needed in the implementation of these newer methods.

A Good Example

Though still in the early stages of development, I do know of one company who is using ML in a way that makes me optimistic. In the fall of 2019, Montage and Shaker International merged to become Modern Hire. In the summer of 2019, I heard the guys at Shaker International present on their interview scoring methods at the International Personnel Assessment Conference (IPAC), which uses an ML method called deep learning (shout out to Isaac Thompson for his work on this method).

Interviewers often rate interviewees on different competencies. Then those ratings are used to inform selection decisions. When done well, rating forms with structured interviews are reliable and accurate tools. Using deep learning, Modern Hire is able to use those interviewer ratings to create an automated evaluation process. So, it takes a method that organizational psychologists often use to evaluate applicants in employee selection, and uses that to teach the ML algorithm what it should be doing. It takes a time consuming process and made it faster to get decisions while applying an accepted organizational psych method. The combination of those things makes this a process that I as a consultant would feel much more comfortable recommending to others. As I mentioned earlier, that has not always been the case when discussing these new methodologies. And no, I am not getting paid or sponsored in anyway by Modern Hire. Again, credit where credit is due. These guys are doing some good work in the arena.

Final Thoughts

New developments and technology are going to keep coming and I am hardly the first one to bring these points up (for example, see Louis Tay’s LinkedIn article). Artificial intelligence and machine learning are the next big thing that people are looking to in advanced analytics. New techniques still need to adhere to the same standards of use by which the typical methods abide. Analyses that cut down the evaluation time, but still results in adverse impact or does not relate to job performance ultimately fall short of becoming a regular in the consultants tool box. With companies like Modern Hire being creative about these new techniques though, we are well on our way to new advancements.

To end, here are a few helpful links related to this article:

Algorithmic Justice
Artificial Intelligence for Talent Assessment
Modern Hire
PWC AI Resources

As always, thanks for reading.

-Brandon
For more expert tips and content, scroll down to subscribe!