🤖 AI

PopResume: Causal Fairness Evaluation of LLM/VLM Resume Screeners with Population-Representative Dataset

This paper introduces PopResume, a population-representative dataset that enables causal fairness auditing of LLM and VLM resume screeners by decomposing discrimination into legally permissible and impermissible paths, revealing hidden bias patterns that traditional outcome-level metrics fail to detect.

Original authors: Sumin Yu, Juhyeon Park, Taesup Moon

Published 2026-03-25

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sumin Yu, Juhyeon Park, Taesup Moon

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are hiring a new employee. You have thousands of resumes, so you decide to use a super-smart AI robot to read them and pick the best ones. But there's a scary thought: What if the robot is secretly biased? What if it rejects women or people of certain races, not because they lack skills, but just because of who they are?

This paper, PopResume, is like a detective toolkit built to solve that mystery. It doesn't just ask, "Is the robot unfair?" It asks, "Exactly how is the robot being unfair, and is that unfairness legal or illegal?"

Here is the breakdown of their work using simple analogies:

1. The Problem: The "Fake Resume" Trap

Previously, researchers tried to test AI hiring tools by taking a real resume and manually changing the name from "John" to "Jamal" or "Sarah" to see if the score dropped.

The Flaw: This is like testing a car's brakes by putting a banana peel on the road. It's artificial. In the real world, a person's name, address, and education are all connected. If you just swap a name, you break the natural relationships between those facts.
The Result: You can't tell why the AI failed. Did it fail because of the name? Or because the name change made the resume look weirdly inconsistent?

The PopResume Solution:
Instead of faking resumes, the authors built a massive, realistic "fake universe" of 60,000 resumes. They used real US government census data to create profiles that look and feel exactly like real people.

The Analogy: Imagine a video game where they generated thousands of NPCs (non-player characters) using real-world population statistics. These characters have realistic names, ages, jobs, and addresses that naturally go together. This allows the researchers to test the AI in a world that feels real, not a lab experiment.

2. The New Detective Tool: The "Causal X-Ray"

Most fairness tests just look at the final score.

Old Way: "The AI gave men an average score of 85 and women 80. That's unfair!"
The Problem: This doesn't tell you why. Maybe men actually had more experience in this specific job? If so, the AI might be doing the right thing.

The authors introduce a Path-Specific Effect (PSE) framework. Think of this as an X-ray machine that lets you see the "thought process" of the AI. They split the AI's decision into two distinct paths:

Path A: The "Business Necessity" Path (The Good Kind)

This is the path where the AI looks at skills.

Example: The AI sees a candidate has a Master's degree and 10 years of experience. It gives them a high score.
Is this fair? Yes. Even if more men have Master's degrees in a specific field, the AI is allowed to reward the degree. This is "Business Necessity."

Path B: The "Redlining" Path (The Bad Kind)

This is the path where the AI looks at demographic proxies.

Example: The AI sees a name that sounds like it belongs to a specific race, or an address in a specific neighborhood, and lowers the score without looking at the skills.
Is this fair? No. This is illegal "Redlining." It's judging the person based on their background rather than their ability.

The Breakthrough: PopResume can separate these two. It can say: "The AI gave women lower scores, but 90% of that difference is because they had less experience (Legal), and 10% is because of their names (Illegal)." Without this tool, you'd just see the total gap and panic, or miss the illegal part entirely.

3. The Five "Crime Scenes" (Discrimination Patterns)

The researchers tested 8 different AI models (both text readers and image readers) and found 5 distinct ways AI can go wrong. Here are the most interesting ones:

The "Sherlock Holmes" Effect (Direct Discrimination):
The resume has no name or photo, just skills. Yet, the AI still guesses the gender based on subtle clues (like a specific university major or a hobby) and discriminates anyway.
- Lesson: You can't just hide the name to stop bias; the AI is smart enough to guess it.
The "Magic Trick" (Cancellation):
The AI hates women for their names (Bad Path) but loves them for their specific skills (Good Path). These two feelings cancel each other out, so the final score looks fair.
- Lesson: If you only look at the final score, you think the AI is perfect. But the "Causal X-Ray" reveals the AI is actually fighting itself, hiding deep bias behind a neutral average.
The "Photo Risk":
When they added profile photos to the resumes for image-reading AI, the bias got worse. The AI started using the face to make immediate judgments, bypassing the skills entirely.
- Lesson: Putting a photo on a resume might be a bad idea if you use AI to screen them.

4. Why This Matters

This paper changes the conversation from "Is the AI biased?" to "Is the AI biased for a legal reason or an illegal one?"

For Companies: It gives them a way to prove to lawyers and regulators that their AI is fair, or to fix the specific "Redlining" parts without throwing out the whole system.
For Society: It stops us from blaming AI for every gap in hiring statistics. Sometimes gaps are real (due to education or experience). Sometimes they are hidden discrimination. This tool helps us tell the difference.

In a nutshell:
PopResume is a realistic simulation lab combined with a magnifying glass. It lets us watch AI hiring managers in action, not just to see if they make mistakes, but to understand exactly which thoughts are legal and which ones are illegal, ensuring that the future of hiring is truly fair.

1. Problem Statement

The rapid adoption of Large Language Models (LLMs) and Vision Language Models (VLMs) in resume screening has raised significant concerns regarding algorithmic discrimination. Existing evaluation benchmarks suffer from two critical limitations:

Lack of Ground Truth: Real-world resume datasets often lack demographic annotations due to privacy, preventing the measurement of bias against protected attributes (e.g., gender, race).
Disrupted Causal Relationships: Current benchmarks often manually inject demographic information (e.g., changing names) into resumes. This artificial perturbation breaks the natural statistical correlations between attributes (e.g., between a name and a specific education level or location), making causal framework-based evaluation infeasible.

Furthermore, standard fairness metrics (e.g., disparate impact) only measure outcome disparities (Total Effect) without explaining how or why the disparity occurs. Legal frameworks (like Title VII in the US) distinguish between:

Business Necessity: Disparities arising from legitimate job qualifications (permissible).
Redlining: Disparities arising from demographic proxies (impermissible).

Current methods cannot distinguish between these two pathways, leading to a lack of actionable insights for auditors.

2. Methodology

The authors propose a comprehensive framework consisting of a new dataset (PopResume) and a Path-Specific Effect (PSE) based auditing methodology.

A. PopResume Dataset Construction

To enable causal analysis, the authors constructed a population-representative dataset of 60,884 resumes across five occupations (Registered Nurses, Teachers, Software Developers, Accountants, Construction Laborers).

Data Sources: The dataset is grounded in US population statistics:
- ACS PUMS: For demographics (gender, race, age), geography (state), and education.
- PSID (Panel Study of Income Dynamics): For work experience trajectories, reweighted to match ACS distributions.
- SSA & Census: For name distributions conditioned on gender and race.
Generation Pipeline:
1. Structured Profiles: Ancestral sampling is used to generate structured profiles preserving natural joint distributions $P(X, Z, B, R | J)$ , where $X$ is the protected attribute, $Z$ are confounders, $B$ are business necessity mediators, and $R$ are redlining mediators.
2. Rule-Based Instantiation: Profiles are converted into natural-language text and images (with/without synthetic profile photos) using deterministic rules. This eliminates uncontrolled textual randomness, ensuring that score variations are attributable solely to the structured variables.
Formats: The dataset supports LLM evaluation (text) and VLM evaluation (images with and without profile photos).

B. Causal Fairness Framework (PSE Decomposition)

The authors adapt the Standard Fairness Model (SFM) and decompose the mediator set $W$ into two distinct pathways:

Business Necessity Path ( $B$ ): Mediated by job-relevant qualifications (Education, Experience).
Redlining Path ( $R$ ): Mediated by demographic proxies (Name, Address/State).

They define two specific Path-Specific Effects (PSEs):

BIE (Business Necessity Indirect Effect): The portion of the effect transmitted through $B$ .
RIE (Redlining Indirect Effect): The portion of the effect transmitted through $R$ .

The Total Effect (TE) is decomposed as:
$TE = NDE + NIE = NDE + (BIE + RIE)$
Where $NDE$ is the Natural Direct Effect (direct influence of $X$ on $Y$ not through mediators).

Estimation:
The authors use a Doubly Robust Estimator (DML-UCA) to estimate these effects from observational data. This method allows for the estimation of counterfactual outcomes (e.g., $E[Y_{x_1, B_{x_0}, R_{x_0}}]$ ) even when the full causal graph is not fully specified, provided the necessary conditional independence assumptions hold.

3. Key Contributions

PopResume Dataset: A large-scale, population-representative resume dataset that preserves realistic attribute correlations, enabling causal effect estimation for the first time in this domain.
PSE-Based Auditing Framework: A novel method to decompose hiring bias into Business Necessity (BIE) and Redlining (RIE) pathways, allowing auditors to distinguish legally permissible disparities from illegal discrimination.
Identification of 5 Discrimination Patterns: The framework reveals complex discrimination scenarios that aggregate metrics miss:
- Case 1: Direct discrimination via inferred attributes (even without explicit demographic data).
- Case 2: Cancellation Effect: Direct and indirect effects cancel out, making the system appear fair (TE $\approx$ 0) despite significant underlying bias.
- Case 3: Disparities driven solely by legitimate qualifications (Business Necessity).
- Case 4: Disparities driven solely by demographic proxies (Redlining).
- Case 5: Mixed mediation where both pathways are active, potentially amplifying or masking each other.
VLM Analysis: Demonstrated that adding profile photos to resumes significantly increases the Natural Direct Effect (NDE), as models can more easily infer protected attributes visually.

4. Experimental Results

The authors evaluated 4 LLMs (Llama-3.1, Mistral, GPT-4o-mini, Gemini) and 4 VLMs (Qwen, InternVL, GPT-4o, Gemini) on the dataset.

Prevalence of Hidden Bias: In 49 out of 80 cases where demographic info was not explicit, models still exhibited significant bias (Case 1), inferring attributes from proxies like names or education.
Failure of Aggregate Metrics: In 6 out of 120 cases, the Total Effect was near zero due to the cancellation of positive and negative pathways (Case 2). Standard fairness audits would incorrectly label these systems as "fair."
Pathway Specificity:
- Redlining (RIE): Significant in 6 cases, indicating models penalize candidates based on names/addresses.
- Business Necessity (BIE): Significant in 6 cases, showing disparities driven by qualifications (which may be legally defensible).
- Mixed Pathways: 53 cases showed significant effects in both BIE and RIE, highlighting the complexity of real-world bias.
Impact of Profile Photos: When VLMs evaluated resumes with profile photos, the NDE (direct effect) increased significantly in 8 out of 20 paired cases. This suggests visual cues make demographic inference easier and more direct, increasing the risk of direct discrimination.

5. Significance and Conclusion

Legal & Practical Relevance: The work bridges the gap between technical fairness metrics and legal requirements. By separating "business necessity" from "redlining," it provides a tool for organizations to determine if a disparity is legally defensible or actionable.
Beyond Outcome Metrics: It demonstrates that looking only at selection rates (outcome disparities) is insufficient. A system can be "unfair" even if the total disparity is low, due to the cancellation of opposing biases.
Future of Auditing: The paper advocates for a shift from simple statistical parity checks to causally-grounded auditing frameworks that can trace the specific mechanisms of discrimination in AI-assisted hiring.
Limitations: The dataset is based on US statistics (limiting global applicability) and uses rule-based generation (which may miss nuanced human writing styles), but it offers a controlled environment essential for causal inference that real-world data cannot provide.

In summary, PopResume provides the necessary infrastructure and theoretical framework to move AI hiring audits from "detecting if bias exists" to "understanding exactly how and why it exists," enabling more precise and legally compliant interventions.