Whose values should a culturally aligned AI represent?

Community labels hid meaningful variation inside each group

The first analysis tested whose values should define the target. Value distributions differed across the four samples, but each community also contained substantial variation across age, gender, region, ethnicity and other measured subgroups. When the researchers restricted coverage to particular subgroups or substituted evidence from another culture, the estimated target shifted. The implication is not that every demographic category causes a particular value. It is that a narrow convenience sample can make one part of a community look like the whole.

The authors also asked large language models to simulate members of the target cultures. Cultural prompting moved model responses closer to the relevant human distribution, yet the simulations compressed the diversity of human judgments and disproportionately resembled some subgroups. A synthetic panel can therefore create the appearance of broad consultation while reproducing the model's own training biases. The paper does not establish that all persona simulation is useless; it shows that it is not a substitute for measuring the intended community and retaining individual-level variation.

For product teams, this changes the governance question. A single 'Japanese values' or 'Singaporean values' setting can be both technically convenient and socially misleading. An alignment target should identify the population, the achieved sample, whose views are underrepresented, how uncertainty is handled and whether a user's expressed preference can override the community-level prior. The authors describe culture as an intermediate reference for situations where little individual evidence is available—not a rule that every member must follow.[1]

Abstract surveys did not reliably predict choices in concrete situations

The second analysis compared abstract value profiles with priorities in specific scenarios. A person may endorse benevolence, security or achievement in a questionnaire, yet rank those values differently when deciding whether to help a colleague, meet a deadline, protect a business or disclose disputed information. The paper reports that abstract profiles only weakly recovered situated priorities, which could reverse across topics. Communities also differed in which fine-grained scenarios they considered culturally important or distinctive.

That matters because an AI assistant rarely answers an abstract values inventory. It responds to a worker weighing layoffs against business survival, a family negotiating responsibility, a student asking about authority or a user seeking health advice. A system trained only on national survey averages may sound culturally informed while applying the wrong trade-off to the actual situation. The dataset's community-guided scenario selection is meant to expose those differences, though it still begins with a candidate pool partly filtered by GPT-5.2 and several existing interaction datasets.

The third analysis examined what signal should be learned. Pairwise preferences are common in model training, but a chosen answer may win because it is clearer, longer, more factual or better formatted—not because its underlying value priority matches the reader. Conversely, two people can endorse the same action for different reasons. In the paper's example, keeping staff may reflect benevolence toward workers or security concerns about operational disruption. The bottom-line answer alone cannot distinguish those rationales.

Re3Culture therefore asks participants to select and rank relevant values before writing the contextual response. An LLM judge rated the response as aligned with the annotated values in 95.3% of paired judgments; swapping the top two values reduced alignment. That supports internal consistency, but it is not independent human validation because a model performed the judging. About 34.9% of scenarios showed a strict ordering of values, while many allowed looser or shared priorities. The study's contribution is a richer supervision record, not proof that one canonical hierarchy exists.[1]

Human grounding can reduce stereotypes only if its limits stay visible

For people using AI in advice, education, work or public services, the practical risk is two-sided. A globally uniform assistant may ignore local norms and relationships. A crudely localised assistant may hard-code a stereotype, pressure minorities to conform or treat a probabilistic community pattern as an instruction. The paper's strongest safeguard is procedural: keep distributions and disagreement visible, gather evidence in realistic contexts, and allow individual preferences and rights to take precedence where they are known.

The sample covers four Asian cultural communities and cannot establish a universal method for every society, diaspora or multilingual setting. Participants were recruited through local vendors and Prolific; vendor-inclusive costs ranged from about $15 per participant-hour in China to $28.50 in Japan and Singapore and included management and quality assurance, not only participant pay. The paper reports detailed demographic distributions, but some cells are small and the age and education profiles differ across communities. Sensitive topics and culturally affiliated samples also require continuing scrutiny of consent, privacy, translation and who benefits from dataset reuse.

The work is a preprint and has not been peer reviewed. Several authors are affiliated with Microsoft Research Asia, while others are from Sungkyunkwan University, Singapore University of Technology and Design, the University of British Columbia, the Institute of Science Tokyo and Renmin University of China. The paper does not test a deployed culturally adapted model, measure whether users receive better advice, or compare harms before and after alignment. Its 95.3% consistency result describes agreement between annotated priorities and responses under an LLM judge, not real-world satisfaction or fairness.

The next evidence should include independent replication, genuinely representative or transparently weighted samples, more cultures and diasporas, human validation of the value-reasoning links and prospective tests in actual products. Evaluations should measure whether cultural adaptation improves relevance without increasing stereotyping, discrimination or pressure to conform. Until then, Re3Culture supports a careful conclusion: cultural alignment begins with accountable human sampling and context, not a flag emoji in a prompt.[1]