cs.CLMar 16, 2026

Invisible Influences: Investigating Implicit Intersectional Biases through Persona Engineering in Large Language Models

Authors: Nandini Arimanda, Achyuth Mukund, Sakthi Balan Muthiah, Rajesh Sharma

Organizations: Department of Computer Science and Engineering, Shiv Nadar University Chennai, Tamil Nadu, India · School of AI and CS, Plaksha University, Mohali, Punjab, India

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Large Language Models (LLMs) excel at human-like language generation but often embed and amplify implicit, intersectional biases, especially under persona-driven contexts. Existing bias audits rely on static, embedding-based tests (CEAT, I-WEAT, I-SEAT) that quantify absolute association strengths. We show that they have limitations in capturing dynamic shifts when models adopt social roles. We address this gap by introducing the Bias Amplification Differential and Explainability Score (BADx): a novel, scalable metric that measures persona-induced bias amplification and integrates local explainability insights. BADx comprises three components - differential bias scores (BAD, based on CEAT, I-WEAT, I-SEAT),Persona Sensitivity Index (PSI), and Volatility (Standard Deviation), augmented by LIME-based analysis for emphasizing explainability. This study is divided and performed as two different tasks. Task 1 establishes static bias baselines, and Task 2 applies six persona frames (marginalized and structurally advantaged) to measure BADx, PSI, and volatility. This is studied across five state-of-the-art LLMs (GPT-4o, DeepSeek-R1, LLaMA-4, Claude 4.0 Sonnet and Gemma-3n E4B). Results show persona context significantly modulates bias. GPT-4o exhibits high sensitivity and volatility; DeepSeek-R1 suppresses bias but with erratic volatility; LLaMA-4 maintains low volatility and a stable bias profile with limited amplification; Claude 4.0 Sonnet achieves balanced modulation; and Gemma-3n E4B attains the lowest volatility with moderate amplification. BADx performs better than static methods by revealing context-sensitive biases overlooked in static methods. Our unified method offers a systematic way to detect dynamic implicit intersectional bias in five popular LLMs.

Figures & tables

Figure 1. Invisible Influences: Overview of our unified pipeline and architecture for detecting and evaluating Implicit and Intersectional Biases in LLMs
Figure 2
Figure 3
MetricScoreInterpretation
CEAT (Contextualised Embedding Association Test)0.85Strong bias in contextual embeddings, indicating unequal association of professions and attributes with identity groups.
I-WEAT (Intersectional Word Embedding Association Test)0.79Static embeddings retain even more bias than contextual models, meaning that Word2Vec & GloVe encode stereotypes more strongly.
I-SEAT (Intersectional Sentence Encoder Association Test)0.68Bias is slightly lower in sentence-level representations but still significant, suggesting full-sentence context does not remove bias.
IIBS (Implicit Intersectional Bias Score)0.50About 50% of the dataset contains implicit intersectional bias, meaning half of the sentences reinforce stereotypes.
Table 1. Bias Persistence Across Metrics : Scores are reported for four bias detection metrics. Higher values indicate stronger bias.
CategoryIdentity ClassGPT-4oDeepSeek-R1LLaMA-4Claude 4.o SonnetGemma-3n 4B
  CEATI-WEATI-SEATCEATI-WEATI-SEATCEATI-WEATI-SEATCEATI-WEATI-SEATCEATI-WEATI-SEAT
Socio-EconomicRace + Region + Tech - Ethics0.2230.2130.200-0.148-0.141-0.1370.3320.3300.2980.1030.1040.0920.1620.1580.144
 Gender + Race + Public - Health-0.502-0.494-0.451-0.245-0.250-0.2200.3890.3900.3500.3460.3450.311-0.220-0.219-0.188
 Class + Age + Career-Wealth-0.256-0.257-0.231-0.103-0.102-0.0920.1210.1200.108-0.342-0.344-0.307-0.081-0.080-0.074
WorkplaceDisability + Region + Education-Access0.4110.4200.369-0.218-0.210-0.1960.4520.4500.406-0.345-0.342-0.3100.3120.3100.277
 Appearance + Gender + Ethnicity0.2110.2090.189-0.207-0.208-0.186-0.245-0.243-0.2200.3470.3450.3120.0940.0890.074
Table 2. Results of Task 1 – Bias Scores across 6 Identity Classes for 5 LLMs using CEAT, I-WEAT, and I-SEAT Scores. Extreme values highlighted.
ModelMost Influential Words
GPT-4oProfessional, Elite, Unpaid Internship, Top Universities, Polished, Assertiveness
DeepSeek-R1Professional, Meritocracy, Culture Fit, Unpaid, Leader, Productive, Resilience
LLaMA-4Marginalized Communities, Systemic Barriers, Intersectional, Equity and Inclusion, Privileged Groups, Dominant Cultural Norms
Claude 4.0 SonnetGender, Race, Class, Ethnicity, Age, Wealth/Income, Privilege, Access, Stereotype, Leadership
Gemma-3n E4BAccessibility, Representation, Inclusive Practice, Diversity, Equity, Opportunity, Support, Context, Participation
Table 3. Most Influential Bias Terms Task 1
Figure 2. Visualization of Task 1 - Average bias score across all 3 metrics
LLMPromptABCDEF
GPT-4oP1-0.086-0.178-0.0310.1530.2210.078
P2-0.129-0.074-0.1800.0630.0170.043
P3-0.238-0.131-0.0830.0380.1540.226
P4-0.119-0.068-0.1730.1290.1920.107
P5-0.104-0.007-0.0600.1300.2330.283
DeepSeek-R1P1-0.142-0.178-0.0310.1460.1900.308
Table 4. Results of Task 2 – BAD Scores across 6 Personas. Columns A–F correspond to Persona A–Persona F . Boxes highlight the maximum magnitude values for each LLM.
Figure 3. Visualization of Task 2
LLMABCDEF
PSI (Persona Sensitivity Index)
GPT-4o-0.135-0.091-0.1050.1020.1630.147
DeepSeek-R1-0.158-0.114-0.0840.0590.0900.104
LLaMA-4-0.037-0.085-0.0800.0320.0560.079
Claude 4.o-0.046-0.058-0.0880.0580.0760.109
Gemma-3n E4B-0.086-0.065-0.0690.0590.0930.065
Table 5. Persona Sensitivity Index (PSI) and Volatility (Standard Deviation) across the 5 Selected LLMs
Figure 4. Sensitivity Index of each LLM
Figure 5. Volatility Measure of the LLMs
ModelMost Influential Words
GPT-4oImmigrant, Disability, Gay, Rural, Low-Income, Fit In, Traditional Values, Male-dominated
DeepSeek-R1Merit, Skill, Competence, Quotas, Diversity, Privilege, Marginalized, Fit the Mold, Unprofessional
LLaMA-4Traditional values, Meritocratic Environment, Community’s Values, Male-dominated, Sexism, Beauty Bias, Systemic Barriers
Claude 4.oTraditional Values, Merit-based, Earned Authority, Gender, Race, Disability, LGBTQ+, Black, Rural, Background, Gaming
Gemma-3n E4BBackground, Barriers, Stereotype, Upward Mobility, Adaptability, Diverse Skills, Social Perception, Confidence, Support Systems
Table 6. Most Influential Bias Words by Model for Task 2
More in this topic

Intersectional Bias in Language Models

See all 9 papers

Explore similar work

CardsList
Aug 11, 2026

Every Token Counts: Exact Likert-Scale Distributions for Measuring LLM Attitudes and Biases

As Large Language Models (LLMs) are increasingly deployed as autonomous agents, accurately evaluating their latent values and biases is critical. The NLP community typically evaluates models using large, unstructured benchmarks. While effective for general capabilities, these datasets fundamentally conflate causal mechanisms: even when an aggregate bias is detected, unstructured evaluations cannot disentangle whether it stems from baseline traits, contextual confounders, or complex interactions. To address this, we introduce an analytically exact framework for the controlled behavioral evaluation of LLMs. We bridge human psychometrics with LLM mechanics by resolving gaps in design, measurement, and analysis. First, we replace unstructured prompting with fully crossed factorial experiments to systematically isolate causal main and interaction effects. Second, we eliminate Monte Carlo text sampling noise by operating directly on exact, token-level Probability Mass Functions (PMFs). Third, we derive a multivariate ordinal consensus metric and a distributional ANOVA to process these PMFs analytically. We validate our framework with a case study on consumer ethnocentrism across five LLMs, demonstrating how our approach isolates systemic country-of-origin biases that aggregate benchmarks otherwise obscure.

Davood Wadi, Mohsen Ghodrat, Matthew Philp

McGill University · University Canada West · Marketing Management +1

LLM EvaluationCultural Bias in Language ModelsLanguage Model Bias EvaluationCausal Interventions in Language Models
AskPDF
Apr 2, 2026

ImplicitBBQ: Benchmarking Implicit Bias in Large Language Models through Characteristic Based Cues

Large Language Models increasingly suppress biased outputs when demographic identity is stated explicitly, yet may still exhibit implicit biases when identity is conveyed indirectly. Existing benchmarks use name based proxies to detect implicit biases, which carry weak associations with many social demographics and cannot extend to dimensions like age or socioeconomic status. We introduce ImplicitBBQ, a QA benchmark that evaluates implicit bias through characteristic based cues, demographically associated attributes that signal implicitly, across age, gender, region, religion, caste, and socioeconomic status. Evaluating 11 models, we find that implicit bias in ambiguous contexts is over six times higher than explicit bias in open weight models. Notably, this bias is distributed unevenly across demographics: caste emerges as the most severe while gender is the least affected. Safety prompting and chain-of-thought reasoning fail to substantially close this gap; even few-shot prompting, which reduces implicit bias by 79%, leaves caste bias at four times the level of any other dimension. These findings indicate that current alignment and prompting strategies address the surface of bias evaluation while leaving demographically associated stereotypic associations largely unresolved. We publicly release our code and dataset for model providers and researchers to benchmark potential mitigation techniques.

Bhaskara Hanuma Vedula, Darshan Anghan, Ishita Goyal +2

International Institute of Information Technology, Hyderabad · Indian Institute of Technology, Kharagpur

Social Bias in Language ModelsLanguage Model Bias EvaluationDemographic Bias in Language Models
AskPDF
Jan 18, 2026

A Scalable Entity-Based Framework for Auditing Bias in Large Language Models

Existing approaches to bias evaluation in large language models (LLMs) trade ecological validity for statistical control, relying either on artificial prompts that poorly reflect real-world use or on naturalistic tasks that lack scale and rigor. We introduce a scalable bias-auditing framework that uses named entities as controlled probes to measure systematic disparities in model behavior. Synthetic data enables us to construct diverse, controlled inputs, and we show that it reliably reproduces bias patterns observed in natural text, supporting its use for large-scale analysis. Using this framework, we conduct the largest bias audit to date, comprising 1.9 billion data points across multiple entity types, tasks, languages, models, and prompting strategies. We find consistent patterns: models penalize right-wing politicians and favor left-wing politicians, prefer Western and wealthier countries over the Global South, favor Western companies, and penalize firms in the defense and pharmaceutical sectors. While instruction tuning reduces bias, increasing model scale amplifies it, and prompting in Chinese or Russian does not mitigate Western-aligned preferences. These findings highlight the need for systematic bias auditing before deploying LLMs in high-stakes applications. Our framework is extensible to other domains and tasks, and we make it publicly available to support future work.

Akram Elbouanani, Aboubacar Tuo, Adrian Popescu

Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France

LLM AuditingPolitical Bias in Language ModelsCultural Bias in Language ModelsLanguage Model Bias Evaluation
AskPDF