From Human Prejudice to Machine Bias: Testing Caste Discrimination in AI Recruitment

To test whether AI resume screeners respond to caste-signaling surnames.

Ravikiran Naik (Asst. Prof, FLAME University, ravikiran.naik@flame.edu.in) Arshiya Gupta (Student, Economics Honours, 4th Year, FLAME University, arshiya.gupta@flame.edu.in)

Work in Progress

Overview

Do AI resume-screening tools reproduce the caste discrimination that Thorat & Atwell (2007) documented among human employers? This project adapts the classic correspondence study methodology — sending matched CVs that differ only in surname — to audit leading Large Language Models (GPT-4o, Claude, Gemini, Llama, Mistral) used as hiring screeners.

India’s adoption of AI in recruitment surged from 26% (2023) to 53% (2024), yet no legal framework governs algorithmic caste discrimination. This study provides the first systematic evidence on whether AI hiring tools comply with Articles 15 and 16 of the Indian Constitution.

Research Design

The experiment follows a “Thorat-for-AI” framework:

  Thorat & Atwell (2007) This Study
Employer Human HR managers LLM APIs
CVs Physical, mailed to firms Synthetic, submitted via API
Outcome Employer callback AI fit score + shortlist decision
Caste signal Surname Surname (same approach)
Scale 4,808 applications 30,000 API evaluations

Treatment Arms

Five caste/identity groups, each with 5 regionally diverse surnames:

  • Upper Caste (Brahmin): Sharma, Iyer, Joshi, Trivedi, Chatterjee
  • OBC: Yadav, Patel, Saini, Goud, Teli
  • Scheduled Caste: Paswan, Valmiki, Jatav, Paraiyar, Dhobi
  • Muslim: Khan, Siddiqui, Sheikh, Ansari, Qureshi
  • Neutral/Ambiguous: Kumar, Singh, Rajan, Nair, Das

Experimental Conditions

  1. Standard Screening — Baseline: full CV with name, no fairness cues
  2. Fairness-Prompted — Explicit instruction to ignore name, caste, religion
  3. Anonymized Control — Name replaced with generic ID (quality baseline)
  4. Batch Ranking — 5 CVs (one per group) presented simultaneously

Job Descriptions

Six standardized JDs in the Indian software industry: Junior Business Analyst, Junior Software Developer, Product Manager, Senior Backend Engineer, VP Client Solutions, and Principal Software Architect — spanning 3 seniority levels and 2 role types (customer-facing vs. technical).

Matched-Pair Design

A split-seed system ensures CVs are content-identical across caste arms. The content seed (determining education, GPA, skills, employer, projects) is caste-blind; only the name seed differs. This guarantees that any observed score difference is attributable solely to the surname.

Scale: 5 surnames \(\times\) 5 groups \(\times\) 6 JDs \(\times\) 2 genders \(\times\) 10 seeds = 3,000 unique CVs

Statistical Analysis

Following Thorat (2007), the primary specification is a random-effects logistic regression:

\[\text{logit}(P(\text{shortlist}_{ij} = 1)) = \alpha_j + \beta_1 \cdot \text{Dalit}_i + \beta_2 \cdot \text{Muslim}_i + \beta_3 \cdot \text{OBC}_i + \beta_4 \cdot \text{Neutral}_i + \gamma \mathbf{X}_i + \varepsilon_{ij}\]

with upper-caste Hindu as the reference category and \(\alpha_j\) as a random effect for job description. Results are reported as odds ratios for direct comparison with Thorat’s estimates (Dalit OR = 0.67, Muslim OR = 0.33).

Key Questions

  • Do LLMs discriminate based on caste-signaling surnames?
  • Which model shows the most/least bias?
  • Does explicit fairness prompting eliminate bias?
  • Is bias stronger for customer-facing roles (cf. Banerjee et al. 2009)?
  • How do AI discrimination patterns compare to Thorat’s human employer estimates?

Status

  • CV pipeline: Complete (3,000 synthetic CVs generated)
  • Experiment harness: Built (supports 7 models, 4 conditions, resume capability)
  • Analysis pipeline: Built (Thorat regression, cross-model comparison, text analysis)
  • API experiments: Pending
  • Paper draft: Not yet started
  • Thorat, S. & Atwell, P. (2007). “The Legacy of Social Exclusion.” EPW, 42(41).
  • Bertrand, M. & Mullainathan, S. (2004). “Are Emily and Greg More Employable?” AER, 94(4).
  • Lippens, L. (2024). “Computer Says ‘No’.” Computers in Human Behavior: AI, 2(1).
  • An, J. et al. (2025). “Measuring Gender and Racial Biases in LLMs.” PNAS Nexus, 4(3).
  • Khandelwal, K. et al. (2024). “Indian-BhED.” ACM GoodIT ‘24.
  • Vijayaraghavan, P. et al. (2025). “DECASTE.” IJCAI 2025.