The Reader Who Was Never There

Vincent van Gogh, "Miners" (1880). Workers descending toward a place no one above ground can see. Public domain, via Wikimedia Commons.

More than 90 percent of employers now use automated software to filter or rank the people who apply to them, according to a 2021 Harvard Business School study, and by 2024 roughly 80 percent of companies reported using AI to review résumés.^[1] Among large employers it is effectively universal: 97.8 percent of Fortune 500 companies run an applicant tracking system, the software layer that ingests, parses, and ranks applications before anyone in the building sees them.^[2]

And when a human does look, they don't look for long. The average recruiter spends about eleven seconds on a résumé that survives the initial filter.^[3] Eleven seconds is not reading. It is a glance to confirm what the software already decided.

Stack those two facts together and you get my nephews' silence. Their applications were parsed, scored, and ranked by a system optimizing for signals they were never told existed. The letters they agonized over (the warmth, the personality, the carefully chosen verb) were, to the first reader, formatting noise. They had written beautifully for an audience that had already left the room.

The rules changed, and nobody sent out the memo. Schools still teach a craft optimized for a hiring world that no longer exists.

Walk into almost any high school or college career center and you will hear advice built for 1995: open with an objective statement, use a distinctive template, show some personality, make it pop. Every one of those instincts is now a liability. A two-column design with a clever sidebar can confuse a parser into scrambling your work history. A graphic header can render your name invisible. "Personality" expressed as prose is exactly the thing the first reader cannot see. We are training a generation to perform for a judge who has been replaced by a turnstile.

There Is No Master Score (and That Matters)

Here is where the panic usually sets in, and where it usually goes wrong.

A viral claim resurfaces every few months: that somewhere out there is a single AI score attached to your name, calculated once and broadcast to every employer, an invisible credit rating for employability that blacklists you everywhere at once. The fear is potent because it feels true. Nothing else seems to explain two hundred rejections that all look identical.

The literal version of that fear is wrong. No one number is stamped on your forehead and sold to the world. The company most often named as the culprit, Pymetrics (acquired by the hiring-software firm Harver in 2022), does not even work that way: it builds trait profiles from short neuroscience-based games rather than scanning résumés.^[4] There is no master score.

But the reassuring opposite, the comforting picture of a fresh and independent judge waiting at every company, turns out to be just as wrong. That is the part almost nobody tells job seekers.

In May 2026, a Stanford-led team led by Rishi Bommasani and Percy Liang published the largest empirical study of algorithmic hiring ever conducted, presented at the ACM Conference on Fairness, Accountability, and Transparency. They analyzed 3.4 million real applicants and roughly 4 million applications screened by a single vendor's models, the same Pymetrics platform.^[5] What they found carries a name borrowed from agriculture: algorithmic monoculture. The same vendor's models are quietly reused across many different employers, so the same applicants are rejected again and again, at rates far higher than chance would predict. Apply to ten positions and there is roughly a 4 percent chance you are screened out of all ten before a human reads a word. To push that risk down to near zero, the researchers calculate, you would need to apply not to ten jobs but to twenty-five.

Sit with what that means for two hundred applications. My nephews were not running two hundred independent experiments. They were drawing, over and over, from the same few decks, shuffled the same few ways. The silence was not two hundred separate verdicts. It was one verdict, echoed.

Regulators are still a step behind this. New York City's Local Law 144 now requires bias audits of hiring tools, and the European Union's AI Act classifies them as high-risk, yet both regimes inspect each employer in isolation, which is precisely the blind spot a cross-employer monoculture slips through.^[6] You cannot wait for the rules to catch up. What you can do is treat the screener as a dialect to learn rather than a lottery to lose: apply more widely than feels reasonable, vary the kinds of roles and employers you approach so you are not drawing from one deck, and pour real effort into the human channels that route around the models entirely. The myth tells you to give up. Learning the grammar tells you where to push instead.

What the Machine Is Optimizing For

Algorithmic literacy starts with a single question you can ask of any automated system: what is this thing optimizing for, and what is it using as a stand-in for me?

When the system is a résumé screener, the answer is alignment. Most systems are not asking "is this person impressive?" They are asking "does this document match this posting?" – measured through keywords, structured fields, parseable work history, and increasingly, semantic similarity between your experience and the job's requirements. The screener is a matching engine, not a talent scout.

That reframing dissolves a lot of bad advice. A single polished "master résumé" sprayed across two hundred postings is almost guaranteed to misalign with most of them. The literate move is the opposite: fewer applications, each one mirrored to the specific language of the specific posting. If the listing says "inventory reconciliation," the résumé that says "inventory reconciliation" aligns; the one that says "kept the stockroom straight" does not, no matter that they describe the same work.

But (and this is the part the keyword-stuffing crowd misses) the systems are getting better at catching crude manipulation. Early tricks like pasting the entire job description in white text, or cramming a skills section with terms you can't back up, increasingly backfire as screeners move from literal keyword matching to semantic analysis that reads context. The skill is not gaming the machine. It is mirroring it: using the real language of the field to describe work you actually did. Speak the machine's dialect honestly, and you are legible. Try to con it, and the next generation of screener flags you.

The Bias in the Mirror

Édouard Manet, "A Bar at the Folies-Bergère" (1882). The reflection in the mirror does not quite match the woman standing before it. Public domain, via Wikimedia Commons.

There is a harder reason algorithmic literacy matters, and it is not about getting ahead. It is about knowing what you are up against.

In 2024, Kyra Wilson, a researcher at the University of Washington, and Aylin Caliskan, a professor there, ran a careful test. They took three open-source large language models, fed them 120 names statistically associated with white and Black men and women, and asked the models to rank otherwise-comparable résumés across more than 500 real job listings, over three million combinations in all.^[7] The models favored résumés with white-associated names 85 percent of the time. Female-associated names were preferred just 11 percent of the time. Résumés with names associated with Black men fared worst of all, disfavored against other candidates nearly 100 percent of the time.^[8]

One year later, the same team found something quieter and more unsettling: when people were shown an AI system's rankings, they tended to mirror its biases in their own subsequent choices.^[9] The machine's prejudice doesn't just filter candidates. It teaches the humans downstream to filter the same way.

This is not a reason to despair, but it is a reason to be clear-eyed. The systems making first contact with your work carry the patterns of the data they were trained on, and those patterns are not neutral. Algorithmic literacy includes knowing that the playing field has a tilt – and that the tilt is precisely why the human-shaped end-runs around the filter still matter so much.

The Hidden Workforce

Gustave Caillebotte, "The Floor Planers" (1875). The labor that keeps a building standing, performed by people the room is built to overlook. Public domain, via Wikimedia Commons.

None of this is hypothetical, and none of it is small. Joseph Fuller, a Harvard Business School professor, led a 2021 study with Accenture estimating that more than 27 million Americans are "hidden workers" in the United States, people capable of doing the jobs they apply for who are nonetheless filtered out, in large part, by the automated systems built to find them.^[10] The classic example is a screener configured to require a four-year degree for a role that plainly doesn't need one: a proxy, standing in for "qualified," that quietly discards millions of capable applicants.

That cost now shows up in the mood of an entire generation. In late 2025, Daniel Chait, the chief executive of the hiring platform Greenhouse, gave the pattern a name in Fortune: an "AI doom loop," in which candidates use AI to mass-produce applications, employers use AI to mass-filter them, and both sides lose faith in the whole exercise.^[11] In Greenhouse's survey of 1,200 U.S. job seekers, nearly half said their trust in hiring had fallen over the previous year, a figure that climbed to 62 percent among Gen Z entry-level workers. Only 8 percent believed the algorithms screening their applications made hiring fairer.^[12] Nineteen percent of organizations using AI in hiring admitted their own tools had screened out qualified people.^[13]

My nephews are two data points inside that statistic. So, probably, is someone you know.

A Literacy Starter Kit

So what do you actually do? If the goal is to turn that diagnosis into something usable, here is the starter kit I gave my nephews – less a list of résumé hacks than a way of seeing.

Read the machine's grammar. Use a clean, single-column, text-based layout. No graphics, no text boxes, no two-column tricks, no critical information buried in headers or images. Standard section labels – Experience, Education, Skills. You are writing a document to be parsed first and admired second. Make the parsing effortless.

Mirror the posting, don't spray a master copy. Tailor each application to the specific listing, using the actual nouns and phrases the employer used. Ten mirrored applications will outperform a hundred generic ones. Align honestly (describe real work in the field's real language) and skip the keyword-stuffing tricks that newer semantic screeners now catch and penalize.

Keep a human in the loop, deliberately. The single most reliable way past an algorithmic filter is to not be in it alone. A referral, a direct message to a hiring manager, a portfolio link, a five-minute conversation at a counter – these are not old-fashioned. In an automated funnel, they are the cheat code. Spend a real share of your effort on the human channels, not just the application portals.

Audit your algorithmic footprint. This is the transferable skill, the one that outlasts any single résumé. Get in the habit of asking, of every system you move through: what is this optimizing for, and what is it using as a proxy for me? Practiced on the job hunt, it becomes the reflex that protects you everywhere a machine reads first.

The Canary Wears a Lanyard

Because the résumé is only the beginning. The same logic (a model making first contact, optimizing for a target you can't see, standing in a proxy for the real you) is spreading into the systems that decide who gets a loan, an apartment, a spot in a program, a match, a reputation. The résumé filter just happens to be the first algorithm most people consciously fight. It is the canary in the mine: a small, early warning about the air in the whole tunnel.

Across the last century, the dividing line in the information economy was access, decided by who had the computer, the connection, the tools. That divide is closing. The new one is subtler and harder to legislate: not who has the machines, but who can read them. Who understands what they want, who can speak to them without being swallowed by them, who can tell the difference between mirroring a system and surrendering to it.

I am going back to my nephews with all of this. Together we are going to rebuild their résumés as plain, parseable text, pick a dozen jobs instead of a hundred and mirror each one to its posting, and put real effort into the human channels a filter cannot see. None of it guarantees a callback. All of it changes the odds.

The work itself never changed. The young men never changed. What changed is that they finally understand who (and what) are reading. That understanding is the literacy of the decade ahead, and almost no one is being taught it.

The first machine to read you shouldn't get to write your ending. Learning its language is how you stay the author of your story.

Disclosure: This article is published by Sage.is, a product of Startr LLC. The views expressed are those of the editorial board.