How Job Websites Use Artificial Intelligence to Filter Resumes (2023)
Exposing the dark side of bias, keywords, and context.
As the job market becomes increasingly competitive, the use of Artificial Intelligence (AI) in the form of Applicant Tracking Systems (ATS) to screen resumes has become increasingly commonplace. Many popular job posting websites such as LinkedIn and Indeed use AI-powered ATS algorithms to filter out or sort candidates and streamline the hiring process. In light of recent hiring freezes in tech and rumours of a recession in the U.S. economy, it’s important for every job candidate to know how they may be affected by AI-based recruiters. Equally important is for human recruiters to recognize the limitations and pitfalls they can run into using AI, and to protect more diverse candidates throughout the hiring process.
The use of AI in recruitment raises ethical and practical concerns about bias, keyword optimization, and the folly of imperfect training data. This article aims to explore the dark side of AI in recruitment from both a technical and non-technical perspective, and provide insight into how job seekers can make their resumes stand out to a robot.
Applicant Tracking Systems
The use of AI to screen resumes on online job postings is a new emerging technology designed to streamline the recruitment process and save employers time and money. These algorithms can quickly analyze resumes, identify key skills and qualifications, and filter out unqualified candidates. Sometimes a single job posting may receive thousands of applicants and this would take a human a much longer time to go through than a robot.
Some Applicant Tracking Systems are simple enough just to screen resumes for keywords. Others are complicated enough to use advanced neural networks trained to find the most successful candidate in a specific field or for a specific position.
According to a survey done by JobScan, 98.8% of Fortune 500 companies use ATS as well as major job websites such as LinkedIn or Indeed. The use of AI-powered ATS is designed to easily eliminate unqualified candidates and prioritize qualified candidates as companies are receiving more job applications online than ever before.
However necessary it may be for employers, the use of AI to judge resumes has significant drawbacks along with its potential advantages:
Biased Decision Making
One of the main concerns with AI-powered recruitment, especially simplistic models, is the potential for biased decision-making. Simplistic rule-based models that are based on keywords or job descriptions may be easily tricked by applicants who tailor their resumes to ATS with keywords and miss qualified candidates who don’t. Meanwhile more advanced algorithms using predictive modelling are often trained on historical data or mock data, and if that data has an incidental bias, the algorithm will may make biased decisions in the hiring process based on irrelevant criteria.
Minority Groups
For example, a modern ATS using Machine Learning is typically trained on historical data of successful candidates. Candidates with certain names or from certain schools were more successful at being hired in the data it is shown so the algorithm will favour those candidates, regardless of their actual qualifications for the job. This can often lead to discrimination against candidates from underrepresented groups and perpetuate systemic biases in hiring especially towards minority applicants. Since we don’t want to base our present hiring on cultural issues we have had in the past, training recruitment AI’s on exclusively historical data of success could effectively reverse the progress an organization has made towards diversity.
One possible implication is that AI-based recruitment could lead to an algorithm favouring white sounding names in a job that previously hired mostly white employees, such as on wall street or favouring women above men in jobs that mostly hire women, such as teaching. Even if we don’t want to bias AI by including factors such as gender or ethnicity, there are many examples differences in resumes from minority groups simply based on having a slightly different background that are impossible to completely isolate an AI from.
For example, a candidate having attended a prestigious but historically black university may be filtered out in the hiring process of a historically white field or company simply because few previous candidates would have attended this university. At the same time completely removing university names from the hiring process wouldn’t necessarily be appropriate as someone with a fashion degree from Parsons in NYC would be a stronger candidate than someone from a less prestigious fashion school applying to a fashion job. It doesn’t just stop at names and university names though. Even something as simple as being a member of the Lacrosse Club versus the Hispanic Honor Society can bias an AI if one of these backgrounds is more likely to be held by the previously successful candidates that it was trained on and one is less likely.
The Solution: Reducing Bias
So do hiring managers have to completely stop using ATS to ensure they are hiring diverse candidates?
Not necessarily — but they do need to make sure they are using systems that value objective attributes rather than simply being a member of a majority group of candidates. One solution is to not allow an ATS to see obvious data that could lead to explicit bias against minorities by pre-processing i.e. removing data containing gender, names and other irrelevant traits. Along with this one should ensure traits with implicit bias are translated into simpler terms more merit-based terms. University names for instance could be scrubbed from resumes before filtering them through an AI and replaced with a ranking based on how prestigious the university is considered for the field. Similarly an ATS could look at number of clubs a candidate has on their resume without learning the actual name of the clubs. Each data point on a resume should be gone through and careful thought given on how to reduce the amount of bias an AI-recruiter can develop towards them during the design of a system.
In addition careful consideration should be given to the diversity of the training data in order to ensure they are representative of as diverse a group of successful candidates as possible in order to prevent a neural network from biasing towards the majority group.
Keyword Selection
One of the key benefits of using AI-powered recruitment tools is that they can disqualify candidates who do not meet specific job requirements. However, this approach can also lead to qualified candidates being disqualified simply because they do not meet the standard educational or job requirements in title, even when their skills and experience are transferable to the job.
For instance, an AI-powered recruitment tool may filter out candidates who do not have a degree in computer science, even if they have relevant experience and hold a slightly different degree in a related field. For example, a candidate may have a degree in Information Technology (IT) or Data Science, which covers many of the similar concepts to a computer science degree, but the AI algorithm may automatically disqualify the candidate because their degree name does not exactly match the job requirements. A very particular neural network may even yield false negatives by looking at historical data of degrees for someone who held a position and because 60% of successful applicants had a degree in “Computer Science” filter out all candidates with degrees of a different name such as Data Science, Computer Engineering etc. because it perceives these candidates as having 40% or less chance of being a successful applicant.
Transferable Skills
This can be problematic for candidates from underrepresented groups who may not have had access to the same educational or career opportunities as their peers or candidates with unique educational backgrounds. If an algorithm is designed to filter out candidates without a specific degree, it may unfairly disadvantage candidates who have equivalent skills and experience but do not hold that exact degree. By only considering candidates who meet very specific criteria, employers miss out on qualified candidates from diverse backgrounds who have skills and experience that are transferable to the job. A recent Harvard Business Study suggests that around 27 million hidden workers exist in the U.S. alone in large part due to Applicant Tracking Systems.
In some cases an atypical candidate could be a better fit to a job yielding a higher profit for the company long-term. Someone with a bachelors in “Robotics and Embedded Systems” might be a much better fit than someone with a “Computer Science” degree for an embedded systems programmer. Robust neural networks for hiring should be tested on how easily they are able to perceive candidates with transferable skills, rather than just how easily they can find a typical successful candidate. Poorly calibrated ATS networks lead to highly qualified atypical applicants being kept out of the labor market. Over time this turns into candidates feeling more pressure to develop general skills that an ATS will approve, such as seeking out more common degrees, participating in more common extracurriculars and attending historically white ivy league universities with high tuition rather than pursuing their unique interests.
The Solution: Keyword Optimization
To combat this issue, job seekers can try to demonstrate how their skills and experience are relevant to the job by using keywords — even if they do not hold the exact degree mentioned in the job requirements or if they suspect it is affecting their ability to get through filters they can experiment with omitting their degree title on their resumes (for example a “Bachelor of Science” vs a “Bachelor of Science in Robotics and Embedded Systems”). It’s important to remember that although both a computer and a human will eventually see your degree, the computer is the first decision maker. Of course its frustrating to have to exclude information that would be relevant to a human-decision maker to an ATS that is not at the level of a human, but the ATS is making the decision to hire you equally as much as your interviewer despite not thinking exactly like a human.
Additionally, employers can consider having HR read over a sample set of rejected resumes at regular intervals and use this to evaluate the success of their AI or any potential bias issues. Ideally employers and developers should work together on any recruitment algorithms making decisions related to the hiring process when possible and make adjustments if possible, but employers do not always have access to the specifics of these algorithms if they are offered through a third party service. In these cases they should attempt to work with third party services that meet their needs for diverse hiring and offers the maximum amount of transparency around their algorithms and data preprocessing.
Recruiters should also be required to conceptually understand the technology behind Applicant Tracking Systems and the potential pitfalls of this technology.
The Training Dataset Problem
When an Artificial Intelligence is trained, it is typically given a training dataset. A recruitment AI that is a neural network would most likely be given a training dataset consisting of resumes and a score of 1 or 0 based on whether the person with this resume would be successfully hired.
The AI would then be asked to filter out resumes under a threshold and keep resumes above a threshold based on how close the scores are to 1. Alternatively the AI could simply be asked to rank all the candidates with scores between 0 and 1, and filter these ranking based on score to send the applicants closest to 1 to the top of a list and the applicants closest to 0 to the bottom. Either option is a version of the AI ranking applicants between 0 and 1, and many ATS seem to use a combination of automatically rejecting and ranking.
But the training dataset isn’t a list of current applicants with a future score of how well they will do because that data isn’t known yet. It is either a mock dataset or a dataset based on historical data. A dataset based on historical data could take records of past applicants resumes and a score of whether or not they were hired. Or it could look at past applicants resumes, whether or not they were hired and how likely they were to stay with the company for longer periods or be promoted within the company. A mock dataset, which is less typical for an ATS, could be made by someone in HR, someone trained in data entry, an interviewer or a programmer looking through resumes and scoring them. A mock dataset is more vulnerable to reflect any accidental bias in the person making it.
Some datasets may give the same weight to candidates and others may give slightly higher or lower weights based off of different markers of success. The goal of finding what dataset might actually yield a successful candidate can be more complicated than just getting hired based on what a company or industry values as successful.
A frightening trend emerging in companies using AI to make decisions is the damage from correlation-based conclusions in a training dataset. Limited AI, such as AI evaluating a candidates chance of success based on their resume, often doesn’t test for causal relationships between data. Causal relationships are relationships in which x precedes y, x is associated with y and that x leading to y can’t be accounted by a third variable. Limited AI systems are often very good at picking up correlational relationships but unlike human decision-makers they don’t grasp rules or causal reasons behind these relationships between data.
If in a training dataset, success in a specific job or field is correlated with an irrelevant piece of data, such as whether or not the applicant was on their high school hockey team or has a common name such as John, the AI may favour candidates with this irrelevant background and disfavor candidates without. But being on a high school hockey team is not more likely to lead to being good at a job any more than having the name John is.
The training dataset problem refers to the issue of biased data being used to train AI algorithms. If the historical data used to train the algorithm is biased, the algorithm will make biased decisions. For example, if the algorithm is trained on data that only includes candidates from certain demographic groups, it will favour those candidates, regardless of their qualifications. This can perpetuate systemic biases in hiring and lead to discrimination against candidates from underrepresented groups.
However, even an ideal dataset that is close to being unbiased can still lead to an AI-based decision maker developing biases towards irrelevant correlational datapoints.
How you can make your resume stand out to a robot
While the use of AI in recruitment raises many ethical concerns about bias and keyword optimization, job seekers can take steps to make their resumes stand out to a robot. Here are some tips:
- Optimize for keywords: Job seekers should tailor their resumes to the job description and include relevant keywords that are likely to be picked up by the AI algorithm. One way to do this is to literally copy and paste the job description into a resume template and then tailor that resume around the keywords from the job description. Applicants should seek to have as many keywords from the job description in their resume as possible, while also keeping their resume professional and relevant enough for a human to view.
- Highlight relevant skills and experience: Job seekers should highlight their relevant skills and experience to ensure that the algorithm recognizes their qualifications.
- Remove irrelevant skills and experience: Job seekers should experiment with removing irrelevant skills and experience to ensure that the algorithm doesn’t disqualify them for these. A common example is someone who graduated with a double major experimenting with removing one of their degrees from the resume if is less relevant to the job they are applying to.
- Don’t exaggerate: Job seekers should avoid exaggerating their skills and experience as this can lead to disqualification if the algorithm determines that the candidate does not meet the job requirements, especially in cases where the applicant may be slightly overqualified for the job.
- Remember a human still has to hire you: Job seekers should keep in mind that their resumes will most likely still be seen by a human decision-maker once they pass an Artificial Intelligence screening and that not all jobs use AI recruiters. This is important because if a resume is completely tailored to an AI but doesn’t stand out as representative and individual to a human eye it won’t get past screenings.
- Use resume filtering tools like VMock and JobScan. These tools use the same technology as many ATS and can help give applicants a rough idea how they are being ranked by these tools.
Conclusion
In conclusion, the use of AI in recruitment has both benefits and drawbacks. While it can streamline the hiring process and save employers time and money, it also raises concerns about bias, keyword optimization, and the training dataset problem. Job seekers can take steps to make their resumes stand out to a robot by optimizing for keywords, highlighting relevant skills and experience, using simple formatting, and avoiding exaggeration. Ultimately, it is up to employers to ensure that their AI algorithms are free from bias and make fair hiring decisions.
References:
