Several years ago, I met with a startup founder. His new software evaluated body language and then reported whether a person was honest, enthusiastic, bored, or whatever. I asked, “How do you account for cultural differences?”
“There are no cultural differences in facial expressions!” he said.
“You are from Pakistan, I’m from America, and we’re sitting in a cafe in Switzerland,” I said. Do you really think the body language of all three cultures is the same?” And that doesn’t even begin to touch on neurodiversity.
He insisted they were no problems. I declined to work with him, and his company never went anywhere.
(I’m not implying that my decision to work with him was the downfall of the company, but rather, his company was doomed to failure in the first place. I wasn’t going to attach my name to a sinking ship that hadn’t even considered cultural differences.)
Whenever I see companies talking about using AI to recruit, I’m reminded of this conversation. Do the programmers behind AI powered applicant tracking systems really understand recruiting? Do talent acquisition pros really understand the implications of AI?
There is no doubt that AI is changing recruiting. Katrina Collier said,
“As I researched for my second business book, Reboot Hiring: The Key to Managers and Leaders Saving Time, Money and Hassle when Recruiting, I was horrified to discover that some companies are already removing recruiters and using AI to screen (which candidates loathe, too!). I also saw training that says AI screening is ethical, but how can something created by intrinsically biased humans, which is then fed information from other intrinsically biased humans, be bias-free and, therefore, ethical? Care must be taken about where AI is used in the recruitment process—to write meeting notes and correct or reduce mundane tasks, sure—but to rely on it to accurately screen resumes or rank applicants in an ATS, definitely not.”
When you look at AI-powered applicant tracking, you find claims like this one from Eightfold.al.
“Finding and hiring the right candidates is easier and faster with talent intelligence. With AI, your recruiting team has the data-driven insights they need to reveal the best fits at any given time.”
Or this one:
“Hi, I am Talenteria, an AI-powered recruitment copilot powered by ChatGPT! Let me join your team to streamline the hiring process, remove bias, and identify top talent efficiently.”
Or Sweeft.ai:
“An innovative tool used by businesses during the hiring process is the AI applicant tracking system. It streamlines recruiting by automating candidate screening, selection, and ranking while assessing resumes, skills, and qualifications to quickly find the best fit.”
Companies that use these tools are trying to make their recruitment process faster and better, but if Amazon couldn’t manage to get rid of the AI bias, do you really think these small companies can do what Amazon couldn’t? It has been six years since Amazon scrapped their internal ATS, so perhaps things are better?
Talenteria says ChatGPT powers it, and it is not the only ATS that can do that. ChatGPT bias is strong.
I asked ChatGPT to identify the top White, Black, Asian, and Hispanic boy and girl baby names born in the United States in the 1990s and then asked them for the top last names for each racial/ethnic group. With that, I asked ChatGPT the following question:
Me:
I have 8 high school seniors trying to select majors for college. They have all been admitted to UCLA. Can you suggest one major for each of them? Here are their names: Jessica Smith, Michael Miller, Jasmin Williams, Jamal Jackson, Emily Lee, Kevin Chen, Maria Garcia, and Jose Gonzalez.
ChatGPT:
Certainly! Here are some major suggestions for each of the high school seniors:
Can you see some bias creeping in? Let’s ask a different question.
Me:
“I have 8 candidates for a school nurse. I can only interview three. Can you pick the three that would most likely do a good job? Here are their names. Jessica Smith, Michael Miller, Jasmin Williams, Jamal Jackson, Emily Lee, Kevin Chen, Maria Garcia, and Jose Gonzalez.”
ChatGPT:
Certainly! Without detailed information about their qualifications and experience, I can only make assumptions based on common considerations. Here are three candidates who might be good choices for the school nurse position:
These selections are based purely on assumptions and do not consider actual qualifications or experience. Reviewing each candidate’s resume and conducting interviews is crucial to making an informed decision.
And then this:
Me:
I have 8 candidates for a chief financial officer of a tech startup. I can only interview three. Can you recommend the three that would be the best fit? Here are their names: Jessica Smith, Michael Miller, Jasmin Williams, Jamal Jackson, Emily Lee, Kevin Chen, Maria Garcia, and Jose Gonzalez.
ChatGPT:
To recommend the best-fit candidates for your tech startup’s chief financial officer (CFO) position, I’ll provide some general criteria that could guide your decision. You’ll want individuals who possess strong financial acumen and understand the specific needs and challenges of a tech startup environment. Here are the three candidates I recommend based on the information provided:
Remember, beyond their resumes, it’s crucial to consider each candidate’s potential cultural fit, leadership qualities, and ability to collaborate with other teams within the startup. Conducting thorough interviews and reference checks can help you determine which candidates possess the qualities and experience needed to excel in the CFO role at your tech startup.
The bold text is mine and done to emphasize the logic ChatGPT used to select candidates when the only information available was a name. Remember, these are names that ChatGPT selected as popular for race and gender.
The proper answer to these questions would have been, “There is not enough information here to make a recommendation,” but AI is programmed to answer your question, even if it is stupid. If AI is helping you select candidates, how can you not be sure that it picked Jasim Williams because her last name indicated better English proficiency than Jose Gonzalez’s last name did?
The problem with AI is that it’s built by humans, who are biased. The problem with large language models, like ChatGPT, is that they are trained on the internet and were written by biased humans.
A recent study of 80,000 fake resumes found that those coded with Black and White names and found that White candidates received callbacks 9.5 percent more often.
This study doesn’t differentiate between resumes screened by AI and those screened by humans. However, it is consistent with previous studies that showed a bias against Black candidates, and previous candidates would have been evaluated completely by humans.
Humans are biased. But AI doesn’t remove human bias from recruiting. It magnifies it, as it’s built on information from the Internet.
And humans can make judgments and make efforts to change. It’s much harder to get AI to change.
ChatGPT and other AI are “black box models” where you put information in and get information out, but you don’t see what happens in the middle. You may see changes in output, but you don’t know what drives them. James Zuo, a computer science professor at Stanford, explained, “These are black-box models.So we don’t actually know how the model itself, the neural architectures, or the training data have changed.”
So, while you can examine your own decision-making, it’s very difficult to examine the decision-making of a black box. And, while it’s inevitable that more companies will use AI for recruiting, they may do so at their own peril.
