How LLMs Simulate Perspective Across Languages
Exploring how LLMs reproduce historical narratives across English and Arabic
With Prof. Heather Ford
I designed and led a cross-lingual prompt experiment to analyze how GPT-4, Gemini, and DeepSeek generate narratives about the Arab Spring across English and Arabic. By treating model outputs as behavioral data, I assessed how cultural framing, language, and temporal cues influence the model’s reconstruction of historical events.
Most work on LLMs evaluates accuracy or surface-level bias. My work examines how models simulate perspective—how they represent collective memory and shift framing based on linguistic and cultural signals in prompts.
Research Question
When prompted about the Arab Spring, do LLMs vary their responses based on the prompter’s country of origin, language, or temporal framing?
Business Impact
Supports Responsible AI, localization, and model evaluation by surfacing representational bias and cross-market inconsistencies in generative output.
What I did (and am doing)
- Designed a multilingual prompt set (English & Arabic) to test cross-cultural variation in responses.
- Generated and analyzed 3,000 model outputs across GPT-4, Gemini, and DeepSeek.
- Applied cultural bias scoring to quantify systematic differences.
- Evaluating cross-model consistency in historical narrative generation.
Why it matters
This study treats LLMs as interactive cultural interfaces. By testing how users in different linguistic contexts might encounter the same system, the research exposes how design choices and data shaping influence perceived credibility, cultural alignment, and user trust. These are key concerns for global UX and AI product teams.