If the question on repeat in your head is “why am I getting rejected from every job,” the answer might be less personal than it feels. A lot of employers screen applicants with software bought from the same small group of vendors. When that happens, a resume that one model scores low gets scored low at the next company too, and the next. Researchers call this algorithmic monoculture.
Two pieces of research landed on this question in 2026. A Stanford working paper studied 4 million applications run through one vendor’s screening algorithms and found a group of people who were recommended for rejection everywhere they applied. Then on September 29, MIT News covered a paper arguing the picture is messier than “one model, rejected everywhere,” and that the harm depends on the details. Both are worth understanding, because each one points to a different fix, and only one of those fixes is in your hands.
How one hiring algorithm can reject you everywhere
Start with the basic mechanism. Most large employers don’t build their own resume screeners. They license them. The MIT News article notes that “a handful of resume screening algorithms are commonly used by many Fortune 500 companies.” The Stanford paper opens with the same premise: many employers screen applicants with algorithms built by the same few vendors.
That changes the math of applying. If 40 companies each had their own human screener with their own quirks, a rejection at one would tell you very little about the other 39. Each would be a fresh draw. When the 40 companies run the same model, the draws stop being independent. Whatever the model dislikes about your resume (a title it doesn’t map well, a gap, a keyword pattern, a format it parses badly) it will dislike every time it sees it.
MIT’s Manish Raghavan put the worry plainly: “as more people use AI and algorithms to get information and make decisions, there is more of a vehicle for this kind of correlation to occur.” Lending already went through this. Credit decisions that used to rest with individual bankers now mostly run on the same FICO-derived scores, which the MIT piece uses as its reference point for what monoculture looks like when it’s finished.
For a job seeker, the practical meaning is simple and uncomfortable. Sending more applications into the same screen doesn’t give you more chances. It gives you the same chance, repeated.
What the Stanford data shows about systemic rejection
The Stanford Digital Economy Lab paper, “Algorithmic Monocultures in Hiring,” by Rishi Bommasani, Sarah Bana, Kathleen A. Creel, Dan Jurafsky and Percy Liang, is unusual because it uses real screening data instead of a simulation. The authors acquired a dataset of 3 million applicants submitting 4 million applications, all screened by algorithms built by one vendor.
The headline finding on individuals: 4% of applicants who applied to 10 positions were recommended for rejection from all 10. The authors say that rate is higher than you’d expect by chance. That’s the systemic rejection pattern in actual data. A small but real slice of people hit the same wall at every door.
The paper also found clear racial disparities. Of all applications submitted by Asian and Black applicants, 14.74% and 25.87% respectively went to positions that adversely impacted those groups under U.S. employment discrimination standards. If you’re in one of those groups and your job application response rate looks worse than your qualifications should allow, that number is relevant to you.
The most telling line comes at the end of the abstract. Because hiring algorithms give the same output for the same input, the researchers could replay what would have happened if each applicant had applied to every position. Their conclusion: applicants “would need to apply widely in order to ensure their applications are considered by a human.”
Sit with that for a second. The fix the data suggests, inside a monoculture, is volume. That’s the same volume game that AI job search tools have made nearly free, which means everyone else is playing it too.
The MIT counterpoint: it depends on the details
The MIT paper, “Algorithmic Monoculture and its Critics” by Brian Hedden and Manish Raghavan, published in Philosophical Perspectives, pushes back on parts of this story. It’s a useful corrective, and it shouldn’t be read as “relax, the problem isn’t real.”
Their argument on systematic exclusion is about the whole market. If every firm uses the same model, the same number of jobs still get filled. Raghavan: “All the jobs get filled and the same number of people have jobs, but the firms are fighting over the same pool of candidates, which actually drives up wages.” Good news if you’re in the pool the model favors. Cold comfort if you’re in the 4% it keeps turning away.
Hedden also took apart the agency objection, the worry that a monoculture strips candidates of any chance to adjust. “If you have a monoculture where you get to revise your resume and resubmit your materials, then this doesn’t hold up,” he said. In other words, a shared model is only a trap if you keep sending it the same input.
The researchers do identify a real cost. They prove mathematically that monoculture tends to create “informational echo chambers that can hinder exploration.” Firms all chase the same profile and stop discovering candidates who might be better. They suggest fixes on the employer side, like adding randomness to a shared platform or combining several firms’ algorithms into an averaged “ensemble” score, which in their simulations sometimes beat a mix of separate algorithms. Hedden says it’s still an open question how practical that ensembling would be.
Raghavan’s summary is the honest one: “A lot of the answers around the promises and pitfalls of algorithmic monoculture are going to be contextual.”
Put the two papers side by side and a usable picture comes out. Monoculture exists. For most applicants it may not change much. For the subset whose resume the model consistently scores low, it’s a wall, and the two remedies on offer are to change what you feed the model or to get a person to look at you before the model does.
Signs a shared screen is behind “why am I getting rejected from every job”
There’s no way to see which model a company uses. The Stanford abstract doesn’t even name its vendor. But there are patterns worth tracking, and a spreadsheet of your last 30 applications will show them faster than guesswork.
Speed is the first one. A decline that lands within hours, or on a weekend, before any recruiter could plausibly have opened the file, is a strong hint that no person was involved.
Then look at the wording. Rejection emails from different companies that share the same phrasing, structure and sometimes sender format are a tell. Different logos, same machine.
Next, the portals. Note where each application actually lived: the URL of the form, the login you had to create, the layout of the questions. If most of your fast rejections cluster on the same kind of portal, you may be hitting the same screening setup repeatedly. Treat this as a clue, not proof. A tracking system and a scoring model are not always the same product.
Compare channels, too. If your hit rate is near zero on portals and better everywhere else (referrals, recruiter outreach, smaller companies that clearly read applications by hand), and those convert while the portals never do, the problem probably isn’t your experience. It’s how a model reads your resume.
Last, check whether the rejections ignore fit. Getting turned down for roles you’re overqualified for, at the same speed as stretch roles, suggests a filter that isn’t reading closely.
None of these signs proves a monoculture. Together, they tell you where your time is being wasted.
How to route around the shared model
Hedden’s point about revising and resubmitting gives you the first move. If the same model keeps turning you away, change the input. Rewrite the resume in plain structure, with your job titles mapped to the titles in the postings you want, and test the new version on a fresh batch of applications. Then watch whether the fast rejections slow down. That’s a cheap experiment, and the data is yours.
The second move matters more, and it gets at why job boards don’t work for anyone stuck in that 4%. A screening model only sees you if you come in through the portal. A hiring manager who has already read a short, specific message from you is a person making the first call, and the shared model never gets a vote on whether you’re worth a look.
Here’s what that looks like in practice:
- Pick roles where you clearly fit. Outreach works best when the case for you is obvious in two lines.
- Find the person who owns the role. Usually that’s the manager the job reports to, not the recruiter. The posting, the team page and LinkedIn titles normally narrow it to one or two people. The guide on how to research a hiring manager before you reach out walks through it.
- Send a short message tied to their actual work. Three to five sentences. What you noticed, what you’ve done that’s relevant, one clear ask.
- Apply through the portal as well. The message doesn’t replace the application. It makes sure a human has seen your name before the model scores it.
An example message (a composite written for this article, not a real one) might read:
Hi Jordan, saw the platform team is hiring a senior data engineer after the migration you wrote about in the spring. I led a similar move at a logistics company, cutting nightly batch time from six hours to forty minutes. I’ve applied through the site too, but wanted you to have the short version. Worth 15 minutes next week?
That message isn’t scored by anything. A person reads it and decides.
Stop feeding the same screen
Back to the question, then. If you’re asking “why am I getting rejected from every job,” part of the answer may be structural: the same few screening vendors sit behind a lot of portals, and the Stanford data shows some applicants get turned away by all of them. MIT’s research is a fair reminder that this isn’t universal and that it depends on the setup. It doesn’t change what to do if the pattern fits you.
Change the input, track the results, and spend more of your effort on the route the model can’t block. A hiring manager who reads your note before the portal scores you is the cleanest way out of a monoculture, because the model never gets the first word.
Reaching out to hiring managers works. The hard part is the research: figuring out who they are, finding something worth saying, and writing a message that doesn’t sound generic. angld.AI handles that pipeline. Paste a job posting, and it identifies the decision maker, researches them, and drafts a personalized outreach message in about 60 seconds.
