ResearchPod Summary
As job seekers increasingly use generative AI to optimize their resumes, candidates with access to premium LLMs gain an unfair advantage, leading to hiring disparities and reduced decision accuracy for employers. This paper models this challenge as a strategic classification game where hiring algorithms must screen candidates who manipulate their resumes stochastically. The authors propose a two-ticket scheme where the hiring algorithm automatically applies an additional LLM manipulation to every incoming resume and evaluates both the submitted and the newly manipulated version to make a final decision.
The researchers demonstrate that traditional hiring processes are inherently biased when candidates have unequal access to LLM tools. By introducing the two-ticket scheme—which treats the original resume and the hirer-manipulated version as two separate 'tickets' to entry—the authors prove that hiring disparities decrease. Furthermore, they extend this to an n-ticket scheme, proving that as the number of manipulations increases, hiring outcomes converge to a group-independent decision, effectively neutralizing the advantage provided by superior individual LLM access. Empirical validation using an open-source resume screening tool confirms that the two-ticket scheme significantly improves true positive rates and reduces fairness disparities between candidate groups.
This work provides a foundational theoretical framework for understanding how generative AI disrupts traditional applicant screening. It offers a practical, low-cost intervention for employers who want to maintain fairness without needing to detect whether an applicant has used AI or which specific model they accessed. The approach allows organizations to leverage AI to level the playing field, ensuring that hiring decisions are based on candidate qualifications rather than their ability to access expensive prompt engineering tools.
Alex: Automated resume screeners might be rewarding the best chatbot, not the best candidate. A new proposal says the fix is for employers to run every resume through a chatbot themselves.
Sam: That sounds backwards. Before we get to why, whose work is this and what kind of paper is it?
Alex: It's from Lee Cohen, Jack Hsieh, Connie Hong and Judy Hanwen Shen at Stanford, posted to arXiv in February 2025. It's a method paper with proofs, plus an experiment on real resumes.
Sam: Plenty of people use AI to edit resumes now. Why should a researcher outside hiring care?
Alex: Because the paper cites a survey where 57 percent of respondents admitted using AI to write their resume. If paid models polish better than free ones, the screener quietly filters for who could afford the better tool. The authors propose a fix that helps fairness and accuracy at once, without knowing who used what. So we'll cover the problem, the one idea behind the fix, and how far the evidence goes.
Sam: Start with the problem. What did they actually observe?
Alex: They took tech-sector resumes and asked a range of models to improve them with a generic prompt. Then they scored them against job postings with an open-source tool that mimics an applicant tracking system, the software that automatically scores and filters resumes. They saw three things. First, edits usually raise scores, but not always; some resumes score lower afterward. Second, premium models raise scores more.
Sam: How much more? Is this a small edge or something that changes who gets through?
Alex: For a product manager posting at DoorDash, GPT-4o, Claude 3.5 Sonnet and DeepSeek-V3 pushed unqualified resumes, from a different occupation, to look indistinguishable from unedited qualified ones. Cheaper models like GPT-3.5 Turbo didn't move scores much on average. The third observation matters most later: running the same model a second time barely changes anything. The first pass standardizes language and structure, and the second pass stays close to it.
Sam: So this is a gaming problem, which my field would call strategic classification. What's different from the classic version?
Right, strategic classification studies decision rules when people adjust their inputs to get a better outcome. Classically, gaming is costly and deterministic, so people only do it when it guarantees success. Here prompting is nearly free, so almost everyone does it. The outcome is random, because model outputs vary. And the employer can't tell whether a resume was edited, or which model did it. Earlier fairness work in hiring usually assumed you knew group membership. This paper assumes you don't.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.
Sam: How do they formalize an AI edit in a way you can prove things about?
Alex: They split resume features into two kinds. Fundamental features are substance, like skills, experience, education. Style features are presentation, like writing quality and vocabulary. An edit keeps the substance and overwrites the style with a fresh draw from a distribution that depends only on the model. So after editing, your original prose barely matters.
Sam: Models invent things, though. A resume that gains skills isn't just a style change.
Alex: They saw exactly that. With a job-specific prompt, GPT-4o copied project management skills from the job ad into a UX designer's resume. So they switched to a prompt saying the candidate would be punished for lying, and to add no new facts. They checked a sample of outputs by hand and found no fabrications, but they say plainly they didn't inspect all of them.
Sam: So the model is: substance fixed, style rerolled. What do they prove about the usual setup?
Alex: The employer sets one score threshold for everyone, aiming to accept as many qualified people as possible with zero unqualified hires. That's maximizing the true positive rate, the share of qualified candidates who get through, with no false positives. Candidates submit whichever version scores higher. If the privileged group's model tends to produce better outputs, a property called stochastic dominance, then qualified privileged candidates pass at least as often as equally qualified unprivileged ones.
Sam: And the fix is where the backwards part comes in.
Alex: Yes, and the third observation is the key. The employer runs every submitted resume through its own model, scores both versions, and accepts if either clears the bar. Two tickets. A resume already polished by a strong model barely changes. An unpolished one gets the boost it never had. The paper's toy example makes it concrete. Imagine scores spread evenly from zero to ten, and everyone at five or above is qualified.
Sam: So with no editing, a threshold of five is perfect.
Alex: Right. Now the privileged group's model adds one point, and the other group has none. To avoid hiring anyone unqualified, the threshold has to rise to six. Every qualified privileged candidate passes, but only 80 percent of qualified unprivileged ones do. Now the employer's model adds a point to unpolished resumes while barely touching polished ones. The two groups' score distributions match again.
Sam: Does that hold as a general result, or only in the toy case?
Alex: They prove that, if the best threshold stays the same under both schemes, two tickets never widen the gap for any resume. It follows that the group gap in true positive rate shrinks, and both groups' rates go up. They also generalize to many tickets, repeated passes by the employer. If the employer's model is at least as strong as the privileged group's, outcomes converge exponentially to a decision that doesn't depend on group at all.
Sam: That "best threshold stays the same" condition is doing a lot of work. When does it hold?
Alex: That's the limitation I'd flag first. Their simple sufficient condition is that the employer's model is no stronger than the privileged group's. But the many-ticket convergence needs it at least as strong. So the guarantees pull in slightly different directions. The authors conclude that the gain is likely greatest with an employer model as strong as possible, but still weaker than the best candidate model, and urge careful evaluation in practice.
Sam: So what happened on real resumes?
Alex: They used 520 resumes from the Djinni recruiting dataset, half UX designers, half project managers. Half were randomly made privileged with GPT-4o, and the employer also used GPT-4o. The other half had no model, GPT-3.5, or GPT-4o-mini. Results were averaged over 500 train-test splits.
Sam: Give me the headline comparison.
Alex: For the UX role, with the unprivileged group using no model, the gap in true positive rate fell from about 27 points to about one. The overall true positive rate rose from about 22 percent to about 38. For the product manager role the gap roughly halved. Accuracy rose modestly in every condition.
Sam: And when both groups had decent tools?
Alex: With GPT-4o-mini, there was no improvement in the size of the gap. Which fits the theory: the fix helps most when access is very unequal.
Sam: How far would you trust this beyond their setup?
Alex: The proofs are clean, given their assumptions. The experiment is narrower. It's one scorer, Resume Matcher, which works on word similarity. The authors note it's the only open-source tracking tool available. It's two job types, one model family, and one simple prompt. They also couldn't hit exactly zero false positives in practice, so they approximated, and thresholds shifted slightly between schemes. They call for audits of real black-box and human-in-the-loop systems.
Sam: So who should read the full paper, and where should they start?
Alex: People working on algorithmic fairness, strategic classification, or hiring technology. Start with the two-ticket section and its guarantees, paying attention to the same-threshold assumption. Then read the empirical validation on the resume dataset, comparing the no-model condition with the GPT-4o-mini one.
Sam: And for everyone else, the one line to keep?
Alex: When candidates can buy polish, give everyone the same polish before you score.
Sam: A cheap fix for a cheap advantage, at least on paper.