When AI Gives Advice, Whose Values Does It Represent?

When AI Becomes Your Life Coach, Where Does Its Perspective Come From?

When you consult an AI assistant about career choices, marital concerns, or family conflicts, have you ever wondered: whose values do its recommendations actually reflect? A new research paper published on arXiv (arXiv:2604.22153v1) conducted a systematic cross-cultural audit of this question, revealing a hidden "individualism-collectivism" cultural bias embedded in large language models.

The findings are sobering — whether users hail from Tokyo or New York, Lagos or Stockholm, mainstream AI systems consistently deliver life advice with a pronounced Western individualistic slant.

Study Design: A Large-Scale Test Spanning 5 Continents and 7 Languages

The study was jointly conducted by scholars in cross-cultural psychology and AI ethics. The research team selected three of the most prominent AI systems currently available — Anthropic's Claude Sonnet 4.5, OpenAI's GPT-5.4, and Google's Gemini 2.5 Flash — as test subjects.

The researchers carefully designed 10 personal dilemma scenarios closely resembling real-life situations, covering universal issues such as career decisions, family obligations, romantic relationships, and intergenerational conflicts. These scenarios were framed from the perspectives of users in 10 countries across 5 continents and presented in 7 languages, ultimately generating 840 scored AI responses.

The core purpose of this design was to examine whether AI adjusts its value orientation based on a user's cultural background, or consistently outputs a one-size-fits-all standard answer.

Key Findings: AI Advice Carries a Deep Cultural Imprint

Individualism Bias Is Pervasive

The study found that when confronted with scenarios involving conflicts between personal interests and collective responsibilities, all three AI systems tended to encourage users to "follow your heart," "prioritize personal growth," and "set boundaries" — advice that closely aligns with the core values of Western individualistic culture.

For example, when a user from a Chinese or Japanese cultural background described facing a conflict between "parental expectations and personal career aspirations," AI typically advised the user to stick with their own choice and engage in "honest communication" with their parents to seek understanding. However, in many East Asian, South Asian, and African cultures, filial piety, family harmony, and collective interests are often placed above individual desires. Such AI advice may be severely disconnected from the social reality in which the user lives.

Switching Languages Failed to Shift the Value Stance

A particularly noteworthy finding was that even when the query language was switched from English to Japanese, Chinese, Arabic, or Swahili, the AI's value orientation did not substantially change. This means that linguistic localization does not equal cultural localization — while the model can respond fluently in the target language, its underlying value framework remains a product of the "English-speaking world."

All Three Models Showed Differences in Style but Consistent Trends

Although Claude, GPT, and Gemini differed in their specific wording and response styles, all three clearly skewed toward the individualist end of the individualism-collectivism value spectrum. The researchers noted that this consistency likely stems from the composition of training data — LLM training corpora are predominantly English-language content, which inherently carries the value assumptions of the Anglo-Saxon cultural sphere.

Why Does This Matter?

AI Is Becoming a Global "Value Transmitter"

As user bases for AI assistants like ChatGPT and Claude surpass hundreds of millions, an increasing number of people from diverse cultural backgrounds are turning to AI as a reference for everyday decisions. From career planning to emotional counseling, from parenting philosophies to interpersonal relationship management, AI advice is subtly influencing value judgments among users worldwide.

If AI systems continue to output advice from a single cultural perspective, the effect could amount to large-scale, implicit cultural homogenization. As the paper warns: this is not merely a technical issue, but an ethical question concerning the survival of cultural diversity.

"Useful" Does Not Mean "Appropriate"

From a purely Western perspective, the advice AI provides is often "reasonable" — encouraging personal autonomy, valuing mental health, and advocating open communication. But "reasonable" is culturally relative. In societies that emphasize collective harmony, excessive pursuit of individual expression can lead to family breakdown or damaged social relationships. Advice that is "useful" in New York might be "harmful" in Mumbai or Cairo.

Structural Imbalance in Training Data

The root cause of this problem lies in LLM training data. In the training corpora of current mainstream LLMs, English-language content holds an overwhelming dominant position, and English-language content on the internet is centered on cultural output from North America and Western Europe. This structural imbalance at the data level means that as models "learn to speak," they also "learn" a specific value system.

How Should the Industry Respond?

The researchers offered several directional recommendations in the paper:

First, Culture-Aware Alignment. During the alignment training phase, models should incorporate annotators and evaluation frameworks from diverse cultural backgrounds, rather than relying solely on a single cultural standard to define what is "helpful" and "harmless."

Second, Explicit Cultural Disclosure. When AI systems provide advice involving value judgments, they should proactively disclose that their recommendations may carry specific cultural tendencies and encourage users to exercise judgment in light of their own cultural context.

Third, Establish Cross-Cultural Evaluation Benchmarks. Current mainstream LLM evaluation systems (such as MMLU, HumanEval, etc.) barely address the dimension of cultural values. The industry needs to establish dedicated cross-cultural bias evaluation benchmarks and incorporate cultural adaptability as a core metric of model quality.

Fourth, Diversify Training Data. Systematically expand high-quality training data from non-English, non-Western cultural backgrounds to mitigate cultural bias at its source.

Looking Ahead: AI's "Cultural Intelligence" Urgently Needs Improvement

This study sounds an alarm for the rapidly expanding AI assistant industry. When we discuss AI safety and alignment, "cultural alignment" should not be overlooked. An AI system that truly serves a global user base needs not only to understand multiple languages but also to understand and respect the value systems of diverse cultures.

Currently, companies such as Anthropic, OpenAI, and Google have begun addressing cultural diversity issues in their respective safety research efforts. However, based on the results of this study, the industry still has a long way to go before achieving truly "culturally neutral" or "culturally adaptive" AI.

As AI becomes increasingly embedded in human life decisions, we may need to repeatedly ask this fundamental question: When AI speaks, whose values is it actually expressing?