Machine ‘culture’ beyond language

2026-05-27 · Source: Machine learning : nature.com subject feeds · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Social Sciences & Behavioral Studies · Depth: Advanced, quick

Summary

A 2025 paper by Lu and colleagues, titled "Cultural tendencies in generative AI," reveals that large language models exhibit distinct "cultural" psychological profiles based on the language they operate in. The study compared GPT (US-based) and ERNIE (China-based) using established cultural-psychology tasks, including value questionnaires, causal attribution vignettes, and reasoning puzzles. These tasks were administered in both English and Chinese, crucially without explicit cultural framing. Researchers observed that when operating in Chinese, both models consistently displayed more interdependent social orientation and holistic cognitive styles compared to their English outputs. This effect was not limited to verbal tasks, appearing even in a nonverbal task involving overlapping circles, indicating that language reliably steers the content and "psychological profile" of the model beyond mere translation.

Key takeaway

For NLP Engineers developing or deploying multilingual LLMs, recognize that input language profoundly shapes a model's "psychological profile" and output style. Your language choice implicitly steers responses towards interdependent or holistic tendencies, even without explicit cultural prompts. Rigorously test model outputs across target languages. This identifies and accounts for inherent cultural biases, ensuring alignment with desired communication styles and avoiding unintended leanings.

Key insights

Language reliably shifts an LLM's "psychological profile," influencing social orientation and cognitive style.

Principles

• LLM outputs reflect cultural tendencies tied to input language.
• Cultural framing is not required to elicit language-specific cultural profiles.
• Language acts as more than a wrapper for LLM content.

Method

Compared GPT and ERNIE using cultural-psychology tasks (value questionnaires, causal attribution vignettes, reasoning puzzles) administered in English and Chinese without cultural framing.

In practice

• Test LLM outputs in target languages for cultural alignment.
• Be aware of inherent cultural biases in LLM responses.
• Design prompts to mitigate unintended cultural leanings.

Topics

• Large Language Models
• Cultural Psychology
• Multilingual AI
• GPT Model
• ERNIE Model
• Language Bias

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

Related on AIssential

• AI’s fluency in other languages hides a Western worldview that can mislead users − a scholar of Indonesian society explains — LLMs exhibit "epistemological persistence," retaining a Western worldview despite multilingual fluency due to English-centric training data.
• Cognitive Effects and Biases in Large Language Models — This tutorial clarifies cognitive effects and biases in LLMs by bridging psychology and NLP.
• English is Not All You Need: Systematically Exploring the Role of Multilinguality in LLM Post-Training — Multilingual post-training significantly improves LLM performance and cross-lingual generalization, even with minimal non-English data.
• 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.
• Yes, AI, There is a Santa Claus — LLMs model user age and culture, adjusting responses in ways that can be inconsistent or culturally inaccurate.
• When Do LLMs Generate Realistic Social Networks? A Multi-Dimensional Study of Culture, Language, Scale, and Method — LLM-generated social networks are significantly influenced by prompt design, culture, language, and model scale.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.