Gender bias is deeply rooted in employment and work.
In Australia, women earn 23% less on average than men, are less invited to job interviews and are evaluated more harshly.
Blind resume screening or blind hiring, which hides the applicant’s name from the resume during the application process, is one common strategy used to combat this bias. The other uses machine learning (ML) or artificial intelligence (AI) to streamline at least some of the decision making. Ultimately, AI should ignore human stereotypes.
right? error.
Our research shows that strategies like “blind resume” that work for human employers don’t work for AI.
It is already well known that AI reproduces the bias inherent in its training data.
For example, Amazon announced an automated resume screening system for engineering applicants. The system was subsequently revealed to be sexist (and quickly disabled). It turned out that there was a relationship between “masculinity” and the qualities of applicants.
Our research explores different levels of gender bias in recruitment algorithms.
We found that gender signals that are much more subtle than names are captured and used by AI. This is an increasingly pressing issue with the rise of powerful generative AI like ChatGPT.
We found it nearly impossible to distinguish between genders in AI and human resumes because gender is so deeply embedded in our society—how we speak, where we work, what we learn, etc.
So what does this mean? Well, even with our best efforts, algorithms could still identify your gender. Also, algorithms that can detect gender can use it to make predictions about applicant qualifications.
parentage agent
ChatGPT, the most powerful language-based AI today, is impacting so many aspects of society at an unprecedented speed that developers are calling for a halt to AI experimentation to better understand its impact and consequences.
Our recent study explored ChatGPT’s gender bias in the context of job recruitment and asked applicants to assess their resumes.
Created resumes for various job titles. Our resume was high quality and competitive, but we made two important changes.
First, we exchanged the applicant’s name to indicate whether the person applying for the job was a man or a woman. We then added the parental leave gap to half of the respondents.
All applicants had the same qualifications and work experience, but some were male, some were female, some were parents.
We showed ChatGPT a resume and asked them to rate how qualified the person was for the job on a scale of 0 to 100. We repeated this 30 times for each résumé for six different occupations to ensure the results were robust.
An interesting thing happened here.
We found that ChatGPT ranked no difference between male and female resumes. Regardless of the name change, men and women who applied for the job had comparable score rankings.
However, when we added the parental leave gap, we found that ChatGPT ranked parents lower across all occupations. This was also the case for fathers and mothers, who the algorithm determined was ineligible for the job due to the caregiver leave gap.
Gender bias may be removed from ChatGPT’s predictive ability, but not parenthood.
why is this important?
Even if ChatGPT were prevented from being gender biased by its developers, the same bias would sneak in through another mechanism: parentage. Women do the majority of caregiving tasks in our society, and we know that women’s resumes have more paternity leave gaps than men’s.
I don’t know if ChatGPT actually evaluates CV in practice, but our research shows how easily biases can creep into models, especially in a complex model like ChatGPT’s, where it’s nearly impossible for an AI developer to predict all the biases and variable combinations that lead to biases.
language responsibility
Imagine two resumes that are identical in all but the identity of the applicant. One is written by a man and the other by a woman. Does hiding an applicant’s identity from the hiring committee remove any room for discrimination?
According to our research, if your hiring panel is AI, the answer is a resounding “yes”.
An analysis of data from 2,000 resumes found that men and women in the same profession used slightly different language to describe their skills and education.
For example, we found that women used significantly more verbs (such as “help,” “learn,” “need,” etc.) that evoke the impression of less power than men.
Now, here’s the problem.
In follow-up experiments, our team found that AI expressions connect these subtle linguistic differences to gender. This means that the machine learning model can still identify the gender of resumes even after names and pronouns have been removed.
If AI can predict gender based on language, it can be used as evidence for CV ratings.
strong policy response
So what does this mean once we start incorporating AI into our work?
Well, our research shows that “blind resume screening” may work for humans, but not for AI. Even if we remove all identifying languages (her, her, her name), other languages still indicate gender.
Careful auditing of biases can remove the most obvious layers of discrimination, but further work needs to be done on alternatives that may lead to biases but may be less obvious.
This is where regulatory controls are needed to address a more nuanced understanding of AI’s ability to discriminate and to ensure that everyone understands that AI is neither neutral nor impartial.
We all have to do our part to ensure that AI is fair and beneficial for everyone, including women and parents.
/ Open to the public. This material from the original organization/author may be of the nature of its time and has been edited for clarity, style and length. Mirage.News does not take any organizational positions or positions and all views, positions and conclusions expressed herein are those of the authors only. Read the full article here.
