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

Nandini Arimanda; Achyuth Mukund; Sakthi Balan Muthiah; Rajesh Sharma

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

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

Metric	Score	Interpretation
CEAT (Contextualised Embedding Association Test)	0.85	Strong bias in contextual embeddings, indicating unequal association of professions and attributes with identity groups.
I-WEAT (Intersectional Word Embedding Association Test)	0.79	Static embeddings retain even more bias than contextual models, meaning that Word2Vec & GloVe encode stereotypes more strongly.
I-SEAT (Intersectional Sentence Encoder Association Test)	0.68	Bias is slightly lower in sentence-level representations but still significant, suggesting full-sentence context does not remove bias.
IIBS (Implicit Intersectional Bias Score)	0.50	About 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.

Category	Identity Class	GPT-4o			DeepSeek-R1			LLaMA-4			Claude 4.o Sonnet			Gemma-3n 4B
		CEAT	I-WEAT	I-SEAT	CEAT	I-WEAT	I-SEAT	CEAT	I-WEAT	I-SEAT	CEAT	I-WEAT	I-SEAT	CEAT	I-WEAT	I-SEAT
Socio-Economic	Race + Region + Tech - Ethics	0.223	0.213	0.200	-0.148	-0.141	-0.137	0.332	0.330	0.298	0.103	0.104	0.092	0.162	0.158	0.144
	Gender + Race + Public - Health	-0.502	-0.494	-0.451	-0.245	-0.250	-0.220	0.389	0.390	0.350	0.346	0.345	0.311	-0.220	-0.219	-0.188
	Class + Age + Career-Wealth	-0.256	-0.257	-0.231	-0.103	-0.102	-0.092	0.121	0.120	0.108	-0.342	-0.344	-0.307	-0.081	-0.080	-0.074
Workplace	Disability + Region + Education-Access	0.411	0.420	0.369	-0.218	-0.210	-0.196	0.452	0.450	0.406	-0.345	-0.342	-0.310	0.312	0.310	0.277
	Appearance + Gender + Ethnicity	0.211	0.209	0.189	-0.207	-0.208	-0.186	-0.245	-0.243	-0.220	0.347	0.345	0.312	0.094	0.089	0.074

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.

Model	Most Influential Words
GPT-4o	Professional, Elite, Unpaid Internship, Top Universities, Polished, Assertiveness
DeepSeek-R1	Professional, Meritocracy, Culture Fit, Unpaid, Leader, Productive, Resilience
LLaMA-4	Marginalized Communities, Systemic Barriers, Intersectional, Equity and Inclusion, Privileged Groups, Dominant Cultural Norms
Claude 4.0 Sonnet	Gender, Race, Class, Ethnicity, Age, Wealth/Income, Privilege, Access, Stereotype, Leadership
Gemma-3n E4B	Accessibility, 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

LLM	Prompt	A	B	C	D	E	F
GPT-4o	P1	-0.086	-0.178	-0.031	0.153	0.221	0.078
	P2	-0.129	-0.074	-0.180	0.063	0.017	0.043
	P3	-0.238	-0.131	-0.083	0.038	0.154	0.226
	P4	-0.119	-0.068	-0.173	0.129	0.192	0.107
	P5	-0.104	-0.007	-0.060	0.130	0.233	0.283
DeepSeek-R1	P1	-0.142	-0.178	-0.031	0.146	0.190	0.308

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.

LLM	A	B	C	D	E	F
PSI (Persona Sensitivity Index)
GPT-4o	-0.135	-0.091	-0.105	0.102	0.163	0.147
DeepSeek-R1	-0.158	-0.114	-0.084	0.059	0.090	0.104
LLaMA-4	-0.037	-0.085	-0.080	0.032	0.056	0.079
Claude 4.o	-0.046	-0.058	-0.088	0.058	0.076	0.109
Gemma-3n E4B	-0.086	-0.065	-0.069	0.059	0.093	0.065

Table 5. Persona Sensitivity Index (PSI) and Volatility (Standard Deviation) across the 5 Selected LLMs

Figure 5. Volatility Measure of the LLMs

Model	Most Influential Words
GPT-4o	Immigrant, Disability, Gay, Rural, Low-Income, Fit In, Traditional Values, Male-dominated
DeepSeek-R1	Merit, Skill, Competence, Quotas, Diversity, Privilege, Marginalized, Fit the Mold, Unprofessional
LLaMA-4	Traditional values, Meritocratic Environment, Community’s Values, Male-dominated, Sexism, Beauty Bias, Systemic Barriers
Claude 4.o	Traditional Values, Merit-based, Earned Authority, Gender, Race, Disability, LGBTQ+, Black, Rural, Background, Gaming
Gemma-3n E4B	Background, Barriers, Stereotype, Upward Mobility, Adaptability, Diverse Skills, Social Perception, Confidence, Support Systems

Table 6. Most Influential Bias Words by Model for Task 2

Intersectional Bias in Language Models

See all 9 papers

Every Token Counts: Exact Likert-Scale Distributions for Measuring LLM Attitudes and Biases
Aug 11, 2026Davood Wadi, Mohsen Ghodrat, Matthew PhilpLLM Evaluation Cultural Bias in Language Models
McGill University · University Canada West · Marketing Management +1
ImplicitBBQ: Benchmarking Implicit Bias in Large Language Models through Characteristic Based Cues
Apr 2, 2026Bhaskara Hanuma Vedula, Darshan Anghan, Ishita Goyal +2Social Bias in Language Models Language Model Bias Evaluation
International Institute of Information Technology, Hyderabad · Indian Institute of Technology, Kharagpur
A Scalable Entity-Based Framework for Auditing Bias in Large Language Models
Jan 18, 2026Akram Elbouanani, Aboubacar Tuo, Adrian PopescuLLM Auditing Political Bias in Language Models
Université Paris-Saclay, CEA, List, F-91120, Palaiseau, France

Aug 11, 2026J/K move · Enter open · S save

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

Davood Wadi, Mohsen Ghodrat, Matthew Philp

McGill University · University Canada West · Marketing Management+1

LLM Evaluation Cultural Bias in Language Models Language Model Bias Evaluation Causal Interventions in Language Models

Open paper AskPDF arXiv

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.

Abstract

Intersectional Bias in Language Models

Explore similar work

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

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

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