It is no news that gender discrimination is prevalent throughout the recruitment and employment process in the workforce. On average, women's salary in Australia is reduced by 23% compared to that of men, with women provided with fewer job opportunities and interviews with recruiters approaching them in a much more stern and critical manner.
Blind hiring and blind CV screening is a process in which prospective employees’ names are concealed from their CVs to ensure that the hiring process is free of any gender bias. Others have resorted to machine learning or artificial intelligence to refine their results to make the selection process easy, believing that AI would be ignorant of those harmful biases.
However, recruiters were wrong. According to research, blind hiring is only a process fit for human recruiters, not AI, which has learned this bias in its training data. Amazon tested this with their engineering applicants by running their CVs through an automatic screening process. The system, however, proved to be sexist as it revealed it had learned the link between 'maleness' and associating that with quality, thus, was promptly discontinued.
As a result, researchers were adamant about finding out the intensity of gender discrimination in algorithms used for recruitment.
They investigated this with ChatGPT by feeding it prompts to evaluate different CVs of prospective employees. They came up with outstanding and ambitious CVs, expanding over many professions. Additionally, they exchanged the names of the applicants to indicate the gender of the individual applying for the position and added a parental leave gap for half of the participants.
The CVs remained consistent with identical job experiences and qualifications, but there were variations in gender, parental status, and non-parental status. They further prompted ChatGPT to rate each individual’s CV based on a system of 0 to 100. To ensure reliability, they replicated this process for six other professions 30 times for each CV.
Results showed that ChatGPT lacked all forms of gender bias and ranked both male and female job applicants similarly, despite changing their names.
However, with the inclusion of the parental leave gap, ChatGPT ranked working parents lower for every profession, applying to both mothers and fathers, rendering them ‘less qualified’ according to the algorithm’s learned response.
This shows that ChatGPT avoided gender discrimination in its CV assessment. However, bias crept in when it came to parenthood. Given that women are more likely to take caregiving leaves in many societies, it follows that CVs of women have a higher likelihood of including parental leave gaps when compared to men.
This opened room for another question: would discrimination still occur if two identical CVs, written by a man and a woman, but hidden in terms of identity?
Research seems to agree.
After investigating 2000 CVs of men and women in the same profession, they seemed to tweak their descriptions slightly regarding their education and skills. For example, women used more verbs that conveyed a sense of low power such as ‘need,’ ‘learn’, or ‘assist’, compared to men.
AI seems to link these slight changes in language by relating them to gender. This proves that machine learning models can recognize gender even after hiding their pronouns and names.
This begs the question: should AI be incorporated into the hiring process?
Researchers concluded that blind hiring is more suitable for human reviewers, but not AI as it continues to pick on detect slight nuances in languages that are influenced by the gender of the CV writer, even when pronouns and names are removed. To tackle this bias, job recruiters should proceed cautiously and carefully examine these intricacies to avoid interference in their assessment. Additionally, implementing regulatory controls is crucial to addressing these biases in AI to ensure a fair and equitable system that benefits all, including parents and women.
Read next: EU's GDPR Fines: Money Talks, but Where Does It Walk?
Blind hiring and blind CV screening is a process in which prospective employees’ names are concealed from their CVs to ensure that the hiring process is free of any gender bias. Others have resorted to machine learning or artificial intelligence to refine their results to make the selection process easy, believing that AI would be ignorant of those harmful biases.
However, recruiters were wrong. According to research, blind hiring is only a process fit for human recruiters, not AI, which has learned this bias in its training data. Amazon tested this with their engineering applicants by running their CVs through an automatic screening process. The system, however, proved to be sexist as it revealed it had learned the link between 'maleness' and associating that with quality, thus, was promptly discontinued.
As a result, researchers were adamant about finding out the intensity of gender discrimination in algorithms used for recruitment.
They investigated this with ChatGPT by feeding it prompts to evaluate different CVs of prospective employees. They came up with outstanding and ambitious CVs, expanding over many professions. Additionally, they exchanged the names of the applicants to indicate the gender of the individual applying for the position and added a parental leave gap for half of the participants.
The CVs remained consistent with identical job experiences and qualifications, but there were variations in gender, parental status, and non-parental status. They further prompted ChatGPT to rate each individual’s CV based on a system of 0 to 100. To ensure reliability, they replicated this process for six other professions 30 times for each CV.
Results showed that ChatGPT lacked all forms of gender bias and ranked both male and female job applicants similarly, despite changing their names.
However, with the inclusion of the parental leave gap, ChatGPT ranked working parents lower for every profession, applying to both mothers and fathers, rendering them ‘less qualified’ according to the algorithm’s learned response.
This shows that ChatGPT avoided gender discrimination in its CV assessment. However, bias crept in when it came to parenthood. Given that women are more likely to take caregiving leaves in many societies, it follows that CVs of women have a higher likelihood of including parental leave gaps when compared to men.
This opened room for another question: would discrimination still occur if two identical CVs, written by a man and a woman, but hidden in terms of identity?
Research seems to agree.
After investigating 2000 CVs of men and women in the same profession, they seemed to tweak their descriptions slightly regarding their education and skills. For example, women used more verbs that conveyed a sense of low power such as ‘need,’ ‘learn’, or ‘assist’, compared to men.
AI seems to link these slight changes in language by relating them to gender. This proves that machine learning models can recognize gender even after hiding their pronouns and names.
This begs the question: should AI be incorporated into the hiring process?
Researchers concluded that blind hiring is more suitable for human reviewers, but not AI as it continues to pick on detect slight nuances in languages that are influenced by the gender of the CV writer, even when pronouns and names are removed. To tackle this bias, job recruiters should proceed cautiously and carefully examine these intricacies to avoid interference in their assessment. Additionally, implementing regulatory controls is crucial to addressing these biases in AI to ensure a fair and equitable system that benefits all, including parents and women.
Read next: EU's GDPR Fines: Money Talks, but Where Does It Walk?
