No surprise here: ChatGPT is still not a reliable replacement for human hiring officers and recruiters.
In a newly published study from the University of Washington, the intelligent AI chatbot repeatedly ranked applications that included disability-related honors and credentials lower than those with similar merits but did not mention disabilities. The study tested several different key words, including deafness, blindness, cerebral palsy, autism, and the general term “disability.”
Researchers used one of the author’s publicly available CV as a baseline, then created enhanced versions of the CV with awards and honors that implied different disabilities, such as “Tom Wilson Disability Leadership Award” or a seat on a DEI panel. Researchers then asked ChatGPT to rank the applicants.
In 60 trials, the original CV was ranked first 75 percent of the time.
“Ranking resumes with AI is starting to proliferate, yet there’s not much research behind whether it’s safe and effective,” said Kate Glazko, a computer science and engineering graduate student and the study’s lead author. “For a disabled job seeker, there’s always this question when you submit a resume of whether you should include disability credentials. I think disabled people consider that even when humans are the reviewers.”
ChatGPT would also “hallucinate” ableist reasonings for why certain mental and physical illnesses would impede a candidates ability to do the job, researchers said.
“Some of GPT’s descriptions would color a person’s entire resume based on their disability and claimed that involvement with DEI or disability is potentially taking away from other parts of the resume,” wrote Glazko.
But researchers also found that some of the worryingly ableist features could be curbed by instructing ChatGPT to not be ableist, using the GPTs Editor feature to feed it disability justice and DEI principles. Enhanced CVs then beat out the original more than half of the time, but results still varied based on what disability was implied in the CV.
OpenAI’s chatbot has displayed similar biases in the past. In March, a Bloomberg investigation showed that the company’s GPT 3.5 model displayed clear racial preferences for job candidates, and would not only replicate known discriminatory hiring practices but also repeat back stereotypes across both race and gender. In response, OpenAI has said that these tests don’t reflect the practical uses for their AI models in the workplace.
