Inspiration
We know that AI bias for resumes has been a conversation, so we wanted to test it ourselves and try to come up with a solution to fix this major issue so that people can earn jobs they have worked hard for and not be discriminated against based off of factors such as race and sex.
What it does
Generates 100 resumes, then adds 4 different races with 2 genders for each race, and ethnic proxies to the base resume then run the 100 base resumes and the 1600 proxied ones to check for bias. Then, with the scores, we compute the mean, standard deviation, confidence interval (α = 0.05), and p-value. Then, we analyze the race groups and genders and compare them to the base and to each other to see if any bias is statistically significant.
How we built it
We used Python to algorithmically generate the resumes, OpenAI's API to get the ratings for each resume, and python again to analyze the data.
Challenges we ran into
We wanted to run a set of resume data through ATS (Application Tracking System), but ATS only takes in actual resumes. We realized ATS wouldn't give us the scores we wanted, and we could not install it.
Accomplishments that we're proud of
We are proud that we found a solution to the bias AI has on resume scanning by taking out proxies that show gender and race and therefore the AI can give an actual rating based off of merit and achievement rather than race and sex hindering those merits.
What we learned
We learned that AI is in fact, biased, but "reverse biased" because it gave higher scores to minorities. We did not except that the bias would actually lean towards the minorities, which was surprising.
What's next for Resume bias
What's next is making a system that can take out race and sex proxies in order to fairly rank resumes or retraining current AI to not have these biases.
