Approximately fourteen seconds after submitting a job application on a business careers page, a candidate’s career path is discreetly determined. Not through a recruiter. Not by a manager of hiring. by a piece of software that scanned the resume, looked for specific keywords ordered in specific ways, computed a score, and either approved the application or stored it in a database that would never be read by a person. This isn’t conjecture. Nowadays, 80% to 90% of big American companies handle the initial phase of almost all job applications. It’s possible that the candidates who were eliminated at this point won’t find out that their resume was turned down before someone saw their name. They are not intended to be informed by the system.
ATSs, or applicant tracking systems, have been around since the late 1990s. In essence, the early iterations were digital filing cabinets where resumes could be kept and then searched. With the help of machine learning, the current generation is far more aggressive. Before any human evaluation takes place, it frequently eliminates 70% to 90% of applicants by actively ranking individuals against the job description. Efficiency is emphasized in ATS companies’ marketing pitches. The fact that efficiency has come at a quantifiable cost to candidates who just so happen to have the incorrect keywords on their resumes is becoming more and more apparent to anyone who has applied for employment in the last three years.
Speaking with professional recruiters gives the impression that the system has created a category of failure mode that didn’t exist ten years ago. The absurdity of the situation was recently described by a senior recruiter at a Fortune 500 organization. The hiring manager’s job description requests that the candidate have prior expertise in “recruiting.” The candidate puts their background under “talent acquisition”—a word that is practically synonymous in the human resources sector—and has twelve years of related experience.
In search of a perfect fit for “recruiting,” the ATS gives the resume a low score. The applicant is turned down. The hiring manager never views the CV of the candidate who would have been the best fit for the job since they were given an automated selection of candidates who just so happened to utilize the exact language the algorithm desired. Everyone thinks the system is functioning. No one is aware that it isn’t.
Among the failure scenarios, the visual formatting issue is arguably the most subtly destructive. Candidates spend hours creating stunning resumes using Canva or other design tools, complete with multi-column layouts, unique artwork, skill-listing sidebars, and crisp typography that appears polished when a human opens the PDF. None of that design is visible to the ATS when it parses the file in its native machine-readable format. It frequently observes missing portions, confused column ordering, and jumbled content. The program finds the résumé that the candidate found impressive to be absurd. Within minutes, the rejection arrives. Never having seen what the parser actually read, the applicant continues to submit resumes in the same manner to the following round of employers, attributing the rejection to other reasons.
The deeper, more legally significant aspect of the problem is the prejudice problem. AI-powered recruiting systems learn on past data, usually the information about successful hires the business has made in the past. Without being specifically instructed to do so, the AI will likely mimic recruiting practices that have historically preferred applicants from particular universities, demographic characteristics, or career paths.
Amazon famously identified this issue with an internal hiring algorithm it had developed, which started penalizing resumes that included the word “women’s”—as in “women’s chess club”—because its training data, derived from a male-dominated technology workforce, had taught it that resumes without such terms were more likely to result in hires. The tool was discarded by Amazon. Many businesses have continued to use similar systems without carrying out equivalent audits since they operate with smaller datasets and less stringent oversight.
In recent years, the EEOC has been more forthright about the legal risks these systems pose. The agency has released guidelines clarifying that firms are still subject to anti-discrimination laws when using AI in hiring. Regardless of whether the tool was created internally or licensed from a third-party vendor, the employer using it is accountable for any disparate results it produces for protected groups, whether by race, gender, age, disability status, or other categories covered by federal law.
The employer and the ATS vendor are named as co-defendants in a number of class action cases that have started to surface in federal court. The legal theory is simple. Because plaintiffs must prove not only that they were rejected but also that the rejection was significantly related to a protected trait, which is sometimes challenging due to the opacity of the underlying algorithms, the practical application is more complicated.
Speaking with job searchers who have applied to dozens or hundreds of postings over the past year, there is a sense that the system has resulted in a certain level of professional demoralization. Within minutes, the rejection notices are delivered. The comments are generic. There is no means for the candidate to find out what specifically led to the rejection, no method to make improvements for the next application, and no way to discern between a rejection due to a parsing issue and one based on valid criteria. There is a substantial cumulative psychological impact. Professionals with talent report feeling as though their careers have stagnated without knowing why. The program is frequently solely to blame, not the applicant. For its part, the software doesn’t provide an explanation.
The regulatory pressure can result in significant transformation. New York City has enacted regulations mandating that companies disclose their use to candidates and carry out impartial bias audits of AI hiring tools. Similar laws have been approved in California. The federal EEOC is still looking into this. Comprehensive AI hiring transparency legislation have been suggested in a number of jurisdictions. If enacted, these laws would oblige employers to give candidates detailed information about the reasons behind their rejection, which would significantly alter how the systems function. It’s unclear if the regulatory framework will keep up with technological advancements before another generation of competent applicants is silently excluded.
Currently, the most useful advice for job searchers is uncomfortable. Make use of a straightforward, one-column resume format that is readable by the parsers. Don’t artificially cram the job description with keywords; instead, match them naturally. Steer clear of visual templates that are impressive to humans but terrible to algorithms. Compose a cover letter that seems authentic rather than artificial intelligence-generated, as recruiters will be able to distinguish between the two when they review the remaining applications. These are not fixes for the underlying issue; rather, they are coping strategies for a dysfunctional system.