At 1:17 on a Tuesday morning, Priya Nair sat at the small pine table in her apartment in Edison, New Jersey, wearing a Rutgers sweatshirt and eating cold poha from a glass container. Her laptop was open to two windows. On the left was ChatGPT. On the right was Workday, the employment portal for a hospital system with 38 open jobs and a password requirement that had already defeated her twice.
Priya had been a clinical coordinator at a dialysis center for seven years. She knew how to calm a diabetic patient who had missed two sessions. She knew which nephrologists returned calls. She knew the quiet panic in the room when a potassium number came back wrong.
None of that appeared in the box.
The box wanted “skills.”
So Priya typed, “Rewrite my resume for a patient operations manager role using keywords from this job description.”
The machine gave her six bullets in nine seconds.
“Drove cross-functional care coordination initiatives to optimize patient throughput.”
Priya looked at the sentence. She had never said “patient throughput” in her life. She copied it anyway.
This is the first small surrender in the modern job search. The applicant stops describing what she has done and begins describing what the machine may reward. The language becomes scented with optimization. Verbs harden. Nouns inflate. A woman who once stayed late to find a Spanish-speaking nurse becomes “a stakeholder-facing operations professional.”
At 1:42, Priya uploaded the file. Workday extracted it, rearranged it, broke one line into three, lost her certification date, and asked her to retype her employment history. At 1:58, she pressed submit.
Somewhere in a cloud server, her ghost went to work.
The ghost resume is written by a machine, read by a machine, judged by a machine, and rejected by a machine. The human who wrote it barely touched it. The human who might have hired for it never saw it. We built a labor market where the first 50,000 decisions are machine-to-machine handshakes, and the humans enter after something breaks.
In 2004, Marianne Bertrand of the University of Chicago and Sendhil Mullainathan of Harvard sent nearly 5,000 fictional resumes to employers in Boston and Chicago. Some carried names such as Emily Walsh and Greg Baker. Others carried names such as Lakisha Washington and Jamal Jones. The resumes were otherwise matched. The white-sounding names received about 50 percent more callbacks. The old system was intimate, and it was unfair.
That study is often used as the origin story for automated hiring. Machines, we were told, would strip away prejudice. A scanner would care about experience, credentials, verbs, dates. It would not hear an accent. It would not admire a golf-school tie. It would not prefer Greg to Jamal because Greg sounded familiar.
Then the scanners learned from us.
Solon Barocas and Andrew Selbst wrote in 2016 that data systems can reproduce discrimination because historical data carries the residue of past choices. A model trained on prior hires learns the habits of prior hiring managers. If a company once promoted men from certain schools, the model may treat those schools as signals. If workers with caregiving gaps were once overlooked, the gap becomes a warning light. The bias gains a new costume. It comes dressed as math.
In a 2021 report on “hidden workers,” Joseph Fuller and Manjari Raman at Harvard Business School found that automated filters routinely screened out candidates for reasons employers later admitted were poor proxies for ability. A six-month gap. A missing degree. A title that failed to match the database. The report estimated that millions of people were left outside consideration, many with the skills employers said they could not find.
The machines did not invent rejection. They made rejection cheaper.
At a logistics company in Ohio, a recruiter named Tom Bender starts his day with coffee in a mug that says WORLD’S OKAYEST DAD and a queue of 1,186 applicants for 14 warehouse supervisor roles. The applicant tracking system has sorted them by fit score. Green means proceed. Yellow means maybe. Red means no.
Tom used to read resumes in stacks. He made notes in blue ink. “Call.” “Odd path, interesting.” “Ask about gap.” Now he reads the top 40. He has meetings at 10, 11, and 1. The system has already moved 900 people into the cold room of no reply.
He trusts it because he has to.
Peter Cappelli, the Wharton professor who has spent decades studying work, argues that employers have made hiring strangely hard for themselves. In “Why Good People Can’t Get Jobs,” he describes companies asking for exact experience, exact software, exact titles, then complaining about shortages. The machine rewards this appetite for precision. A manager wants “three years of Salesforce.” The system searches for “Salesforce.” A candidate who ran the same process in HubSpot may vanish.
Human judgment has a gift. It can see cousins. It can hear that “I organized the night shift” and “I managed labor allocation” may be the same sentence spoken by two social classes.
Machines prefer twins.
Candidates learn this quickly. On Reddit forums and TikTok clips, strangers teach each other to feed job descriptions into ChatGPT, to mirror language, to add invisible keyword blocks in white font, to rename ordinary tasks. The resume becomes less a record than a lure. The applicant is fishing for a bot.
Employers answer with stronger bots. The old arms race was between people who inflated and people who checked references. The new one is between text generators and text classifiers, each sharpening itself against the other. The document in the middle grows more perfect and less true.
This is the hollowing out. Not the arrival of machines alone. The thinning of contact. Hiring once contained awkwardness, gossip, instinct, prejudice, mercy, surprise. The new process removes much of the bad friction and much of the good friction at the same time. It produces clean dashboards from messy lives.
At 8:06 that morning, Priya’s application received its first status update. “Under review.” No one at the hospital had opened the file. The review was a score, a rank, a set of extracted nouns. By lunch, the status changed again.
“Application no longer under consideration.”
Priya was in the dialysis center then, helping an 82-year-old man named Mr. Castillo find the Mets game on the waiting room television. Her phone buzzed in her scrub pocket. She read the sentence, locked the screen, and returned to the remote.
At the pine table that night, the ghost resume still sat in her downloads folder. It was polished, fluent, and empty in exactly the way the system had requested. It had traveled farther than Priya had. It had met another machine in a place no human could enter. Then it came home carrying a verdict.