AI-Generated Resumes
Consider the phrases below.
- Resolve advanced technology issues and maintain public computing environments; provide IT support to students and faculty over the phone and in person
- Developed an android app using java that supports location sharing, instant messaging between users and tracing user paths.
- Worked on web and MySQL for better customer experience.
- Deploying applications in weblogic/jboss servers, configuring data sources on the netsuite ERP financials team.
- Optimized load times through caching, pagination, compression and data encryption algorithms under professor chien wang’s atmospheric research.
At first glance, these seem like ordinary bullet points from an early-stage technical resume of, perhaps, a high school or college student.
Three out of the five points above were artificially generated. The other two were selected from real resumes. Can you identify the real statements?
As a Computer Science student applying to internships, it’s incredibly frustrating to apply to tens (hundreds?) of internship positions, only to be rejected – or not to hear back – from most. On the flipside, recruiters at both large and small companies recieve thousands of applications for SWE internship positions – and, as a result, are forced to skim through most resumes incredibly fast, or filter out resumes based on patterns and keywords using recruiting software.
In this project, I ask the question – if a recruiter were faced with adversarial, intelligently-generated resumes, would they be able to tell the difference between resumes written by humans and resumes written by computers?
I came across Markovify – a simple Markov Chain sentence generator – and decided to try spoofing resume bullet points like the ones above. My training dataset was scraped from Jumpstart [aside: see this post for a small accompanying data analysis portion of this project]. After pre-processing & training on some 75,000 bullet points, I spoofed about 3,000 generated bullet points. I then spun up a quick iOS App to simulate the environment of a hypothetical recruiter (disclaimer: I have no idea what actually happens in the recruiting room, but I assume something relatively close to this sort of speed-reading happens).
Check out the finished project below! The iOS App also on GitHub.