Why AI Keeps Rejecting Your Resume, and Why It Is Probably Not Your Resume

You have sent out eighty applications and heard nothing. Somebody has told you the problem is your resume: add keywords, use a simpler template, mirror the job description, avoid tables.

Some of that advice is real. But there is a peer-reviewed audit of how these systems actually rank people, and it suggests something the career-advice industry has no commercial reason to tell you: a large part of the rejection is a property of the ranking system, not of your resume.

Here is what the research shows, stated precisely, including the parts that do not support the strongest version of the claim.

What the study actually did

The paper is "Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval" by Kyra Wilson and Aylin Caliskan, published at the AAAI/ACM Conference on AI, Ethics and Society in 2024.

The setup matters. Rather than asking a chatbot to judge candidates, they modelled what production systems typically do: document retrieval. Resumes and job descriptions are turned into embeddings, and candidates are ranked by how close their resume sits to the job description in that space. They tested Massive Text Embedding models this way across nine occupations, using over 500 real resumes and over 500 job descriptions, then varied only the names attached to otherwise comparable resumes. That produced roughly 40,000 paired comparisons.

If you have read the embeddings explainer in this series, you already know the mechanism. The system is not reading your resume the way a person does. It is measuring distance between two points in a learned space, and everything in the text, including your name, contributes to where your point lands.

The numbers

Across those comparisons:

• White-associated names were preferred 85.1% of the time. Black-associated names, 8.6%.
• Male-associated names were preferred 51.9% of the time, female-associated names 11.1%.
• Disadvantage compounded at intersections of race and gender, rather than simply adding up.

The resumes were comparable. The names were not. That gap is not a candidate-quality signal, and no amount of reformatting changes it, because the thing being penalised is not the format.

What the study does not show, said plainly

Precision matters here more than outrage, and this is where I part company with how these findings usually get shared.

It tested embedding-based retrievers, not every commercial ATS. Applicant tracking systems vary enormously. Many are keyword filters, some are hand-screened, some use models like the ones tested. You cannot conclude from this paper that every employer's system behaves this way. You can conclude that a common and growing architecture does.

It is a 2024 result. Models change. That cuts both ways: it does not automatically hold for a 2026 system, and it does not automatically stop holding either.

Be careful with the vendor numbers circulating in 2026. You will see confident claims of the "candidates X% less likely to advance, based on an audit of 33,000 jobs" variety. Several of those are self-published by hiring platforms with a product to sell and have not been independently verified. I am deliberately not repeating them. One peer-reviewed audit with a public methodology and public code is worth more than five vendor blog posts, and Wilson and Caliskan published their code.

Why this connects to work I have published

My own research sits next to this. I have been working on what I call artificial frictional unemployment: the delay and dead weight introduced into labour markets by automated intermediaries, where qualified people and real vacancies fail to meet because the matching layer between them is lossy.

Classical frictional unemployment is the ordinary time it takes to find a job. What is new is that a screening system can manufacture friction at scale, silently, and attribute the outcome to the candidate. The Wilson and Caliskan audit is one measured instance of exactly that mechanism, with a number attached.

That framing changes what you do next. If the friction is in the funnel, the winning move is not to polish your input to the funnel. It is to go around it.

What to actually do

The formatting advice that is real: use a plain, parseable layout, avoid tables and text-in-images for anything that matters, use standard section headings, and use the actual words from the job description where they honestly apply to you. This is real because parsers genuinely fail on exotic layouts. It is table stakes, not a strategy.

The cargo cult: keyword stuffing, white-text tricks, cramming the job description into your resume. Retrieval systems compare meaning, not keyword counts, so stuffing mostly produces a document that reads badly to the human who eventually sees it. Do not do this.

The actual strategy is to reduce how many times you pass through an automated ranker at all. Referrals, direct contact with the person who owns the problem, and public evidence of your work. A repository, a write-up, a small system you built and shipped. This is the same argument as the 93% piece: the AI-titled front door is the most contested and most gated route in the building. Being the person a hiring manager already knows can do the job skips the ranker entirely.

And if you build hiring systems: audit yours. The methodology is public and the code is open. Not knowing is now a choice.

What would make me wrong

I am leaning on one study, and one study is one study, however well-conducted. It uses name-based proxies for race and gender, which is a standard audit technique but a proxy nonetheless, and it measures ranking behaviour in a simulated pipeline rather than hiring outcomes at real firms. It also does not tell us what any particular employer does. The honest claim is narrow: a common architecture, measured carefully, showed large name-linked disparities. The dishonest version, which I am avoiding, is "AI hiring is rigged everywhere, and that is why you did not get the job." I do not know why you did not get that job, and neither does anyone selling you a resume template.

Key takeaways

1. A peer-reviewed audit found embedding-based resume rankers preferred white-associated names 85.1% to 8.6%, and male-associated names 51.9% to 11.1%, with compounding at intersections.
2. The mechanism is retrieval, ranking by distance in embedding space, not reading your resume as a person would.
3. It tested embedding retrievers, not every commercial system. Do not overclaim, and treat vendor-published audit stats with suspicion.
4. Basic parseable formatting is real. Keyword stuffing is cargo cult.
5. The strategy is to route around the funnel: referrals, direct contact, and public evidence of shipped work.

If you have been through this recently: how many of your applications went through an automated ranker, and how many came from someone who already knew your work?

---

Sources: Kyra Wilson and Aylin Caliskan, "Gender, Race, and Intersectional Bias in Resume Screening via Language Model Retrieval", Proceedings of the AAAI/ACM Conference on AI, Ethics and Society (AIES), 2024; preprint at arXiv

.20371; code at github.com/kyrawilson/Resume-Screening-Bias. Coverage and discussion via Brookings. My related work on artificial frictional unemployment in automated recruitment systems: arXiv

.14534.