LLM Cultural Alignment Varies by Prompt Framing in Value Surveys

A new study on arXiv tests whether LLM behavior conditioned on human identity—via personalization, role-play, or forecasting—produces interchangeable cultural alignment. Using 101 questions from the World Values Survey (WVS) across 13 language-country slices, researchers evaluated GPT-5.4, Claude Sonnet 4.6, Gemini 2.5 Flash, and Qwen3-235B. Over 21,008 model-response rows, prompt framing emerged as a first-order determinant: country cues often shifted answers substantially, but not always toward matched human distributions. Third-person forecasting yielded the strongest directional alignment for three of four models, while personalization and role-play were weaker or less stable. Alignment gains concentrated on salient value dimensions. The findings suggest that how an LLM is prompted to consider identity critically affects its value alignment, with implications for deploying AI in culturally sensitive contexts.