Artificial intelligence in Human Resources is the greatest, right? It can screen thousands of applications in nanoseconds and narrow the field just to the types of people with whom you've had success in the past, right? It won't know whether an applicant is a man or a woman, good-looking or homely, or white, Black, or Latino, so it will protect you against discrimination claims.
We-ell, AI can be great, but it isn't perfect. If you're using AI to perform hiring functions -- and especially if you're using it for other HR functions, such as promotion decisions, performance management, or discipline and discharge -- you'll need to be careful.
Selecting "who's been good in the past." One AI problem is that the algorithms are often set up to select applicants with traits associated with employees who have been good for the employer in the past. That makes perfect sense. Until you think about it for a minute. Before 1964, it was legal to discriminate based on race, sex, national origin, religion, color, age, and disability (and probably more). Mandatory retirement at age 65 was normal. Disability discrimination didn't become illegal for employers who were not federal contractors until 1992. As a result, until relatively recently, the workforce was made up predominantly of white males.
AI algorithms that look at "who's been good in the past" may still skew heavily toward white male applicants because white males have dominated the U.S. workforce for so long. And it may not be enough to simply remove race, sex, and age from the algorithm, because sometimes the algorithm can read between the lines and figure out that a particular candidate is in the "wrong" demographic based on other available information. For example, if I majored in Women's Studies in college, the algorithm is probably going to assume that I am a female. If I have a lengthy work history, the algorithm may assume that I am older.
AI doesn't always know the difference between correlation and causation. This is closely related to my last point. Just because a company had great success with white male employees for many years (in other words, white maleness is correlated with success), the AI may "believe" that being a white male causes one to be a good employee. This can obviously present problems from an EEO standpoint.
Compared with accounting and other fields, Human Resources is full of "gray areas," which AI doesn't always handle very well. How does an algorithm decide what's "fair," or determine the "optics" of an employment decision? Someday it may be able to do this, but we aren't there yet.
Who's liable if the AI discriminates? Look in the mirror. Let's say you purchase AI from a vendor who appears to be qualified. Then the AI screens out a class action-full of applicants based on their race. Can you sue the AI vendor? We don't know that yet. Can the class members or the EEOC come after your company? We do know the answer to that -- of course they can.