Somewhere right now, a person is pasting the entire text of a job listing into the bottom of their résumé and turning the font white. The idea, passed around in forums and career TikTok like a folk remedy, is that the screening software will read the keywords and the human recruiter will never see them. Whether it works is almost beside the point — plenty of applicant tracking systems strip formatting or flag the trick, and career coaches warn it can get you blacklisted. What matters is what the ritual reveals: millions of people now understand that the first reader of their working life is not a person, and they are adapting their behavior accordingly, largely on superstition, because no one will tell them the actual rules.
This is the part of automated hiring that gets lost in the familiar complaints about robots rejecting résumés. The deeper shift is not that software ranks candidates. It is that software decides who is ever seen at all, using criteria the candidate cannot know, cannot appeal, and in most cases cannot even confirm were applied. The applicant must be perfectly legible to the machine. The machine owes the applicant nothing — not a reason, not a score, not a correction. That asymmetry, more than any single biased algorithm, is what is actually new about looking for a job.
The legibility machine
The folklore exists because the opacity is real. Job seekers trade tips about which file formats parse cleanly, whether columns confuse the parser, whether a gap year needs to be disguised as “consulting.” Some of this advice is useful; much of it is the employment equivalent of knocking on wood. But the underlying instinct is sound. A system that screens for continuous employment will quietly penalize anyone who took time off for caregiving, illness, or a pandemic layoff. A system trained on the profiles of past successful hires will tend to reproduce the shape of those hires. A system that can’t parse a non-standard career — the freelancer, the career-changer, the immigrant whose experience doesn’t map onto domestic job titles — doesn’t reject those people so much as fail to perceive them.
Researchers at Stanford’s Institute for Human-Centered AI recently got a rare look inside this black box, following roughly 3.4 million people through 4 million applications screened by a single third-party vendor. Applying the EEOC’s four-fifths rule — the standard benchmark for adverse impact — they found that more than a quarter of Black applicants and 15 percent of Asian applicants had applied to positions where the system discriminated against their group; had recommendation rates been equal, about 40,000 more of their applications would have advanced. The finding that should unsettle every employer using these tools is how the bias hid: pooled across all positions, the numbers looked acceptable. Only position-by-position analysis exposed it. Fairness, it turns out, can be laundered by aggregation.
The same study documented something stranger. Ten percent of repeat applicants — people who applied four times — were rejected from every single position, far more often than chance would predict. A comparison dataset of applications to Fortune 500 firms showed no such pattern. Whatever the vendor’s system was doing, it wasn’t evaluating each application fresh; something about a person appeared to follow them from rejection to rejection, a scarlet letter stored in a database they don’t know exists.
Screened out before the questions start
For disabled applicants, the machine’s illegibility runs in both directions. Roughly one in four American adults — about 61 million people — has a disability, and their unemployment rate has run about double that of everyone else. A 2025 CUNY Law Review analysis by A. Ghahremani catalogues how the assessment layer of hiring stacks the deck further: timed video interviews that penalize processing differences, gamified tests incompatible with screen readers, platforms whose accommodation processes are buried or nonexistent. When the Center for Democracy & Technology screened major assessment platforms for accessibility, it found barriers across the board — and found that some vendors don’t let applicants review their own results at all. The ADA’s requirements of reasonable accommodation and an interactive process don’t evaporate because the interview is conducted by software, but the software often behaves as if they do. There is no one to ask for extra time when the door is a web form.
The machine on the stand
The regulatory response is arriving, slowly and unevenly. New York City’s Local Law 144, enforced since July 2023, is the most concrete example: employers using automated employment decision tools must have the tool independently audited for bias within a year of use, publish information about the audit, and notify candidates at least ten business days before the tool is used on them. It is a genuinely novel requirement — and a telling one. The law mandates that an audit happen and that notice be given. It does not give the rejected applicant the right to know their score, to see the criteria, or to contest the outcome. Even the most aggressive hiring-algorithm law in the country mostly guarantees that someone, somewhere, checked the machine — not that the machine explains itself to you.
The more interesting fight is over whether the dossier itself is the problem. In a lawsuit reported in January, applicants sued the screening company Eightfold AI — which says its dataset spans a million job titles, a million skills, and the profiles of more than a billion people, and which scores candidates from one to five without telling them — arguing that these hiring scores should carry the same obligations as credit reports under the Fair Credit Reporting Act. The analogy is pointed. We decided decades ago that a company compiling files on people’s financial lives had to let those people see the file and dispute its errors. The algorithmic hiring industry has built something functionally similar — a shadow record of employability, following workers from application to application — while insisting it’s just software.
That’s the reframe worth sitting with. The debate about AI in hiring is usually framed as a question of accuracy: are the machines biased, and can we make them fairer? But the white-text ritualists have stumbled onto the truer picture. Hiring has quietly acquired a credit-bureau architecture without credit-bureau rules — scores you can’t see, errors you can’t correct, judgments you can’t appeal. The applicant is audited. The auditor is not. Until that reverses, the résumé hacks will keep circulating, because superstition is what people invent when the system refuses to publish its rules.