The Hiring System You Are Applying Into Is Not What You Think It Is
The cold-application surface is no longer a market of independent decisions. It is a single decision, repeated.
Stanford published the largest empirical study of AI hiring algorithms ever conducted this week. Four million job applications, 3.4 million applicants, 156 employers, 1,700 job postings, eleven industry sectors. All screened by a single vendor. The paper is titled Algorithmic Monocultures in Hiring and it will be presented at the ACM Conference on Fairness, Accountability, and Transparency next month.
The study has two findings. The first is what the coverage has focused on, and it is real and consequential: the screening system the researchers studied showed racial bias significant enough to flag under the EEOC's four-fifths rule. Twenty-six percent of Black applicants and fifteen percent of Asian applicants applied to positions where the algorithm discriminated against their racial group. That finding stands on its own and is not what this piece is about. It is being covered well elsewhere, and the legal, regulatory, and ethical implications of it deserve their own attention from people whose work is in those frames.
This piece is about the second finding. The one that is less likely to make headlines and more likely to change how anyone running a serious career move ought to think about the search they are running.
Ninety percent of U.S. employers now use AI screening tools to sort and rank job applicants. Most of them rely on the same few third-party vendors. The Stanford researchers studied one such vendor across 156 employers, and found something that the candidate-side conversation has not really begun to absorb.
A single vendor's algorithm, deployed across many employers, produces correlated outcomes across those employers. A candidate rejected by one company is more likely than chance would predict to be rejected by every other company using the same vendor. The researchers call this systemic rejection. It is not a metaphor. It is a measured statistical pattern in the dataset they studied.
In the case of the vendor the researchers examined, candidate assessment scores are stored and reused for up to 330 days, following the same applicant across applications they did not know were connected. The applicant believes they are submitting twenty-five independent applications. The vendor sees one candidate, one score, twenty-five lookups.
The math the researchers ran on this is the part worth sitting with. Among applicants who applied to ten positions screened by this vendor, four percent were rejected by every single one, a rate statistically higher than what chance would predict if each employer were making independent decisions. The researchers calculated that to be reasonably certain of producing even one recommendation to advance, an applicant would need to apply to at least twenty-five different roles. Twenty-five applications. One recommendation. Maybe.
That is the structural shape of the cold-application surface in 2026. Not "the market is competitive." Not "rejection is part of the process." Something more specific: a single algorithmic decision, replicated across employers who think they are running independent processes, returning the same answer about the same candidate across applications the candidate cannot tell are connected.
There is a particular instinct that takes hold when a job search stalls. Applications are not converting. Phone screens are not converting to onsites. The instinct is to send more. More applications, more recruiters contacted, more time at the surface that is producing the rejection.
The instinct is wrong. It was wrong before the Stanford research was published, and the wrongness is now empirically grounded.
A market of independent decisions responds to volume. If each application were a separate decision by a separate decision-maker reading a separate set of signals, then sending twenty more applications would mean twenty more independent chances. The probability calculus would favor volume. The frustrating outcome (three offers from five hundred applications) would still be evidence that the next five hundred applications might produce three more offers, and the search-as-numbers-game frame would hold.
A market of correlated decisions does not respond to volume the same way. If twenty applications go through the same screening algorithm, the algorithm produces twenty correlated outputs about the same candidate. Sending twenty more applications, still through the same screening surface, produces twenty more correlated outputs. The candidate is not running twenty independent attempts. The candidate is running the same attempt twenty times, against a system that has already seen them, decided what to make of them, and stored the decision for the next three hundred and thirty days.
The cold-application surface is no longer a market in the conversational sense. It is closer to a single gate, replicated. Volume against a single gate does not increase the number of independent shots. It runs the same rejection more times, with the same vendor reading the same signals about the same candidate.
This is not a marginal observation. It is the structural fact about the surface most senior candidates are spending most of their search energy on. The candidates who report sending three hundred applications without offers are not running a slow search. They are running a structurally pre-determined search whose outcome was largely settled by the time the first algorithmic assessment was completed.
The candidate-side response to this finding is not to optimize for the algorithm. The candidates who try this (better resume keywords, better assessment-game performance, better tailoring) are responding to the symptom and missing the structure. Optimizing for an algorithmic monoculture means trying to game a system that has already seen you, made its assessment, and stored it for a year. The optimization is fighting last year's decision against this year's application.
The strategically coherent response is to route around the surface entirely. To run a search whose conversion pathway does not depend on the algorithm's recommendation in the first place. This is not a new idea. It is what senior candidates who actually get hired into senior roles have been doing for as long as senior hiring has existed. What is new is that the Stanford research now empirically grounds why the alternative is not just stylistically preferable. It is the only path that scales.
Routing around the algorithmic surface means a few specific things in practice.
Thesis. A candidate who can finish the sentence "I am the operator who does [specific thing] for [specific company stage and shape]" compresses the search from hundreds of postings to a much smaller list of specific companies and people. The thesis is not a marketing exercise. It is the operational filter that determines whether the search has a center.
Targets. A target list of fifteen named companies, researched in depth, evaluated against the thesis, and committed to working as a campaign. Not two hundred postings auto-emailed to your inbox. The target list is finite, updated weekly, and runs on a timeline you control rather than one the employers' posting cadence dictates.
Warm paths. For each of the fifteen companies, where are the people you already know who know people inside? Where are the second-degree connections through advisors, former colleagues, classmates, or the LinkedIn network? Which warm paths can you open inside the next ninety days, against which you can build the relationship before any role is posted? The conversation that produces a senior hire is rarely the one that begins with an application. It is the one that has been quietly developing for months before a role exists.
Phases. Each target sits in a known phase (worth a look, strong match, stretch, in conversation, active interview, offer territory) and the candidate knows what phase every target is in and what the next move is for each one. The phases are visible. The pipeline is auditable. The campaign produces signal about itself every week.
Signal review. Every target evaluated against the week's new signals: funding rounds, hires, departures, public commitments, market movements. The signal review changes the target list. It changes which targets are moving and which are stalled. It changes the phase progression. The review takes thirty minutes a week. It is the difference between a campaign that compounds and a search that drifts.
Operational continuity. The campaign runs while the candidate is doing the rest of their life. Most days are not campaign days. Most weeks have one or two campaign hours. The continuity is the discipline of not dropping the campaign when the rest of life takes the foreground.
This is not faster than cold applications. It is slower in the early weeks and compounds faster in the later ones. The first month is positioning work: thesis articulation, target compression, warm-path mapping. The visible output is low and the underlying work is dense. The candidates who skip this and revert to volume produce activity but not progress. The candidates who do the upstream work produce, in the later weeks, conversation density that the cold-application surface structurally cannot.
Three months of this kind of campaign reliably outperforms twelve months of applications run against the screening surface. Not because the campaign candidate is more qualified, but because they are running a search whose outcome is not pre-determined by an algorithmic monoculture they cannot see.
It would be inaccurate to write about the Stanford findings without naming the regulatory context. The EU AI Act designates hiring algorithms as high-risk AI systems by default; compliance requirements take effect August 2, 2026, within weeks. New York City passed Local Law 144 in 2021, the first U.S. legislation targeting algorithmic hiring directly, though the researchers noted that the existing government guidance under that law appears to instruct auditors to pool data across positions and employers, exactly the aggregation method that masks the disparities the Stanford paper documents.
The honest read on the regulatory layer is that it is real, it is moving, and it is slower than the candidate-side problem. By the time enforcement of the EU AI Act produces auditable changes in how vendors operate, three or four cohorts of senior candidates will have run their entire job searches inside the system the regulation is attempting to govern. The candidate cannot afford to wait for the regulation. The regulation is part of a different timeline running on a different clock.
This is what makes the candidate-side response so important: it is the only response the candidate controls. The candidate cannot fix the algorithmic monoculture. The candidate cannot require the vendor to audit per-position rather than per-vendor. The candidate cannot litigate adverse impact under EEOC standards as a private matter. What the candidate can do is run a search whose conversion pathway does not require any of those problems to be solved first.
The most useful response to the Stanford research, for someone running a serious career move right now, is not a better resume or a better answer to assessment-game prompts. It is a change in what category the candidate believes they are operating in.
The category most candidates are still using, job search, names a transaction. The category that names the work is campaign, and the work has a structure that does not include the screening surface as its primary path. Once the category shifts, the operational decisions follow. The target list compresses. The warm-path map opens. The phases become visible. The signal review becomes a weekly discipline. The search slows down and starts moving.
The Stanford research did not invent this argument. It made it harder to dispute on empirical grounds.
For a candidate three hundred applications into a stalled search, the most expensive line item on the search is not the next application. It is the time spent running the wrong system, against a structural fact about that system that the candidate could not see from the inside.
The fact is now public. The campaign is the work that responds to it.
