Original paper:
https://arxiv.org/pdf/2112.08910.pdfThe entire premise of the research relies on the fact that the resume samples were well matched: in other words, the quality of male candidates was equal to that of female counterparts. In the paper, it says
Specifically, we perform 1-1 matching without replacement such that for each male resume, we find a female resume that is within 2 years of experience, has the same degree, field of study, and has a resume similarity score (i.e. cosine similarity of resume vector representations) of at least 0.7.
I am not sure if this method is sufficient to create a match dataset. In other words, it is entirely possible that the (legitimate) experience or quality differences in resumes may act as the signal for the ML algorithm.
It would be interesting to see descriptive analytics on male and female applicants. IIRC men and women have differences in self-selection for precisely things like this. It is plausible (or even probable) that such difference in self-selection crept into the dataset despite the research's resume matching process.
(As a clarification, this is a critique on the research methodology, and I am not claiming one gender is less qualified than the other in general.)