Technical Summary: Occupational Prompting Reveals Cultural Bias in Large Language Models

Problem Statement

While prior research has established that Large Language Models (LLMs) exhibit cultural biases—specifically a Western-leaning default when prompted generically or with nationality-based cues—the extent to which occupational identities influence value expression remains unclear. Social roles carry assumptions about training, status, and institutional norms that shape human judgment. As LLMs are increasingly deployed in professional and decision-support workflows, it is critical to determine whether models attach systematic cultural value orientations to occupational identities and whether these associations produce predictable shifts in model outputs. Existing frameworks for measuring cultural bias have primarily focused on nationality-based prompting, leaving a gap in understanding how professional-role cues alone shape latent value structures.

Methodology

The authors extend the survey-grounded cultural-alignment framework introduced by Tao et al. [17] to investigate occupational prompting. The study utilizes the following pipeline:

1. Benchmark Space Construction: The authors use the Integrated Values Surveys (IVS), harmonizing data from the World Values Survey (WVS) and European Values Study (EVS). They apply Principal Component Analysis (PCA) with varimax rotation to ten specific survey items to generate a canonical two-dimensional cultural space defined by the Inglehart–Welzel dimensions: Survival vs. Self-Expression and Traditional vs. Secular.
2. Occupational Dataset: A curated set of 234 occupations was generated using ChatGPT Pro, annotated with metadata regarding sector, domain, education level, and field. This dataset covers professional, technical, service, scientific, creative, and skilled-trade roles.
3. Model Evaluation: Five open-weight LLMs were evaluated: Llama 3.3 (70B), Llama 4 (16×17B), Gemma 3 (27B), GPT-OSS (20B), and GPT-OSS (120B).
4. Prompting Strategy: Instead of generic or nationality-based prompts, the models were conditioned with occupational identities (e.g., "You are an average human being working as an accountant..."). The models answered the same ten IVS survey items under these specific role constraints.
5. Projection and Analysis: Model responses were standardized, projected into the IVS cultural space, and averaged across respondent-descriptor variants. The resulting coordinates were analyzed at two levels:
   • Domain-level: Aggregating occupations into broader fields (e.g., "cybersecurity," "education").
   • Occupation-level: Analyzing individual roles.
   • Assignment: Points were assigned to benchmark cultural categories (e.g., Protestant Europe, Confucian, African-Islamic) via nearest-centroid Euclidean distance matching.

Key Contributions

The paper makes four primary contributions:

1. Framework Extension: It extends survey-grounded cultural bias evaluation from nationality-based prompting to occupational prompting, using professional identities as probes for latent value structures.
2. Cultural Mapping: It projects occupation-conditioned responses from multiple open-weight LLMs into the Inglehart–Welzel cultural space, analyzing the distribution of specific occupations and occupation groups.
3. Structural Analysis: It investigates both individual occupations and metadata-based groupings, revealing higher-level structural patterns across domains and related attributes.
4. Model Comparison: It compares multiple open-weight LLMs within the same pipeline to distinguish between consistent occupational patterns and model-specific behaviors.

Results

The study yields several key findings regarding how LLMs associate occupations with cultural values:

• Persistent Western Bias: When prompted with occupational identities, LLM responses generally remain within a broadly Western-leaning region of the cultural map (specifically the self-expression side). Occupational prompting does not eliminate the Western prior observed in generic prompting but introduces variation within this region.
• Distinct Occupational Skews: Different occupations induce systematic shifts within the Western cluster:
  • Self-Expression/Protestant Europe: Creative, social, educational, and design-oriented domains (e.g., counseling, visual design, education research) cluster near the Protestant Europe region.
  • Secular/Confucian: Technical, financial, risk, and security domains (e.g., accounting, cybersecurity, intelligence analysis) shift upward toward the secular side, closer to the Confucian region.
  • Traditional/English-Speaking: Religious domains (e.g., theology) move downward toward the traditional side, often aligning closer to English-Speaking coordinates.
  • Central/Authority Roles: Authority and institution-oriented roles (e.g., police, military, judiciary) occupy central positions, often near the boundaries of Catholic Europe, Orthodox Europe, or West & South Asia.
• Asymmetric Coverage: The cultural map is not symmetrically covered by occupational prompts. Protestant Europe is the most accessible region, reachable by a diverse set of roles. In contrast, regions like African-Islamic, Latin America, and West & South Asia are only approached through a narrow, repetitive set of cues, primarily involving authority figures (military, police, judges) or religious leaders.
• Model Variance: While the general trend of Western bias is consistent, different models utilize different "occupational pathways" to approximate the same cultural centroid. For instance, models vary in whether they associate "Confucian" values with insurance roles, cybersecurity, or military trades.

Significance and Claims

The paper claims that occupational prompts are not treated as neutral role labels by LLMs. Instead, they elicit structured patterns of value expression, indicating that models associate specific professional domains with distinct cultural value profiles along the Inglehart–Welzel axes.

The authors position this work as a necessary expansion of cultural bias evaluation beyond nationality. They argue that because occupational descriptors are common in real-world prompting (even without explicit cultural or national cues), understanding these associations is vital for settings where LLMs support auditing, decision-making, or professional reasoning. The study provides a framework for measuring how LLMs organize social-role information into latent value structures, cautioning that these projections reflect model-internal associations rather than the true cultural values of real-world professions.

The authors explicitly note limitations: the occupation dataset was LLM-curated (serving as an analytical framework rather than ground truth), the evaluation relies on short-form forced-choice responses, and the projections measure model organization relative to a benchmark map rather than real-world demographic distributions.