Occupational Prompting Reveals Cultural Bias in Large Language Models

2026-02-24 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

A study by Eren et al. investigates how large language models (LLMs) associate occupational identities with cultural value patterns, extending prior nationality-based cultural prompting. Using a survey-grounded evaluation pipeline with questions from the Integrated Values Surveys, the researchers projected responses from five open-weight LLMs (Llama 3.3 (70B), Llama 4 (16x17B), Gemma 3 (27B), GPT-OSS (20B), and GPT-OSS (120B)) into the two-dimensional Inglehart–Welzel cultural space. The findings reveal that while occupation-conditioned responses generally remain within a Western-leaning region, specific occupations introduce distinct "occupational skews." For instance, technical and financial roles shift towards more secular values, whereas creative and educational roles align with higher self-expression. This indicates that LLMs do not treat occupational prompts as neutral, but rather elicit structured value patterns, with some cultural regions being less accessible through occupational cues.

Key takeaway

For AI Scientists and Machine Learning Engineers developing or deploying LLMs, understanding occupational prompting is crucial. Your models likely embed cultural biases tied to professional roles, even if not explicitly prompted for nationality. You should evaluate how occupational cues influence value expression in your LLM outputs, especially in decision-support or analytical applications, to mitigate unintended biases and ensure appropriate model behavior.

Key insights

Occupational prompts elicit structured cultural value patterns in LLMs, even within a Western-leaning default.

Principles

• LLMs associate occupations with distinct cultural value profiles.
• Cultural bias in LLMs extends beyond nationality to professional roles.
• Occupational prompts are not neutral role labels.

Method

The study adapted a survey-grounded cultural-bias framework, replacing nationality-based prompting with occupational prompting for open-weight LLMs. Responses to IVS questions were projected into the Inglehart–Welzel cultural space using PCA.

In practice

• Evaluate LLM outputs for occupational value biases.
• Consider occupational context in prompt design.
• Map LLM responses to established cultural frameworks.

Topics

• Large Language Models
• Cultural Bias
• Occupational Prompting
• Inglehart–Welzel Cultural Space
• Value Alignment
• AI Ethics

Code references

• meta-llama/llama3
• meta-llama/llama-models

Best for: Research Scientist, AI Engineer, NLP Engineer, AI Scientist, AI Ethicist, Machine Learning Engineer

Related on AIssential

• Machine ‘culture’ beyond language — Language reliably shifts an LLM's "psychological profile," influencing social orientation and cognitive style.
• Cognitive Effects and Biases in Large Language Models — This tutorial clarifies cognitive effects and biases in LLMs by bridging psychology and NLP.
• The Context Dilemma: Prompting vs. RAG vs. Fine-Tuning — Prompting, RAG, and fine-tuning are distinct LLM techniques for managing context and grounding, not interchangeable solutions.
• Yes, AI, There is a Santa Claus — LLMs model user age and culture, adjusting responses in ways that can be inconsistent or culturally inaccurate.
• CulturALL: Benchmarking Multilingual and Multicultural Competence of LLMs on Grounded Tasks — CulturALL benchmarks LLMs' multilingual, multicultural, and grounded reasoning on real-world tasks, revealing significant performance gaps.
• Divination by Prompt: LLM-Mediated Xuanxue on Chinese Social Media — LLMs facilitate a new form of divination, Xuanxue, on Chinese social media, driven by user curiosity and anxiety.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.