서브메뉴
검색
Cultural Variability and Bias in Online Social Interactions and Large Language Models
Cultural Variability and Bias in Online Social Interactions and Large Language Models
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105224
- ISBN
- 9798291566541
- DDC
- 004
- 저자명
- Seth, Agrima.
- 서명/저자
- Cultural Variability and Bias in Online Social Interactions and Large Language Models
- 발행사항
- [Sl] : University of Michigan, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 133 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-03, Section: B.
- 주기사항
- Advisor: Toyama, Kentaro.
- 학위논문주기
- Thesis (Ph.D.)--University of Michigan, 2025.
- 초록/해제
- 요약Despite their intended global usage, most technologies are designed and developed within narrow cultural frames, reflecting the values and assumptions of their often Western developers. Thus, when deployed across cultures without consideration of different cultural viewpoints, the resulting cultural misalignment fails to serve diverse cultural contexts. This cultural misalignment actively reinforces cultural logics and assumptions that could inadvertently perpetuate existing inequalities. This thesis examines the intersection of culture and computing through three studies that focus on understanding the cultural biases encoded in large language models and the impact of culture on user behavior in online platforms. To assess cultural biases in LLMs, the first study evaluates how well the cultural information encoded in four leading publicly available LLMs aligns with the shared cultural understanding of members from one Global South culture, India. The second study introduces a prompt-based framework to analyze the extent of biases encoded in LLMs, specifically examining underrepresentation and erasure through the task of generating responses to culturally sensitive questions focused on social dimensions such as religion and caste. Finally, the third study explores how cultural values drive user interactions on social media platforms like Snapchat, which deploy uniform features globally, by analyzing friendship network structures and content consumption patterns among users across 73 countries. My first study finds a significant lack of cultural information encoded in these models, with performance varying substantially across subcultures (states) within the country based on the socio-economic status of the region and the localized nature of the knowledge. While accuracy was high for some subcultures, model performance was worse for regions with lower socio-economic status, and performance declined as knowledge became more culturally localized. My second study, an audit of caste and religious identity representation in LLM outputs, reveals that model defaults overwhelmingly favor socially dominant groups regardless of their numerical representation, and that light prompt-based interventions prove inconsistent in mitigating these biases. These two studies demonstrate that despite user feedback requesting corrections and diverse outputs, LLMs consistently favor information from dominant groups and globally popular cultural elements, and they do not adapt their generation to marginalized or locally specific perspectives. This reveals that while information about subcultures exists within the models, it doesn't surface readily, highlighting that the issue extends beyond missing data and includes problems with algorithmic generation. Finally, my third study shows that even on platforms (such as Snapchat) with uniform affordances across all users, cultural values drive significant differences in user behavior, such as friendship network structures and content consumption patterns (Stories). Specifically, compared to collectivism, individualism and high relational mobility lead to the formation of larger and less egocentric friendship networks and negatively moderate the effect of tie strength on content engagement. Collectively, these studies reveal that current algorithmic technologies inadequately serve diverse cultural contexts, demonstrating the need for a conscious effort to understand and incorporate cultural diversity into the data used for training and the design of algorithms that operate on this data to surface culturally diverse information. Finally, through its examination of cultural biases in LLM and cross-cultural user behaviors on social media platforms, this dissertation contributes to the growing field of culturally aware computing by advancing our understanding of LLM cultural alignment and bias, revealing patterns in cross-cultural digital behavior, and providing suggestions for frameworks that allow for developing more inclusive computational systems.
- 일반주제명
- Computer science
- 일반주제명
- Information science
- 일반주제명
- Information technology
- 키워드
- Responsible AI
- 기타저자
- University of Michigan Information
- 기본자료저록
- Dissertations Abstracts International. 87-03B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017359848
■00520260202105224
■006m o d
■007cr#unu||||||||
■020 ▼a9798291566541
■035 ▼a(MiAaPQ)AAI32271834
■035 ▼a(MiAaPQ)umichrackham006210
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a004
■1001 ▼aSeth, Agrima.
■24510▼aCultural Variability and Bias in Online Social Interactions and Large Language Models
■260 ▼a[Sl]▼bUniversity of Michigan▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a133 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-03, Section: B.
■500 ▼aAdvisor: Toyama, Kentaro.
■5021 ▼aThesis (Ph.D.)--University of Michigan, 2025.
■520 ▼aDespite their intended global usage, most technologies are designed and developed within narrow cultural frames, reflecting the values and assumptions of their often Western developers. Thus, when deployed across cultures without consideration of different cultural viewpoints, the resulting cultural misalignment fails to serve diverse cultural contexts. This cultural misalignment actively reinforces cultural logics and assumptions that could inadvertently perpetuate existing inequalities. This thesis examines the intersection of culture and computing through three studies that focus on understanding the cultural biases encoded in large language models and the impact of culture on user behavior in online platforms. To assess cultural biases in LLMs, the first study evaluates how well the cultural information encoded in four leading publicly available LLMs aligns with the shared cultural understanding of members from one Global South culture, India. The second study introduces a prompt-based framework to analyze the extent of biases encoded in LLMs, specifically examining underrepresentation and erasure through the task of generating responses to culturally sensitive questions focused on social dimensions such as religion and caste. Finally, the third study explores how cultural values drive user interactions on social media platforms like Snapchat, which deploy uniform features globally, by analyzing friendship network structures and content consumption patterns among users across 73 countries. My first study finds a significant lack of cultural information encoded in these models, with performance varying substantially across subcultures (states) within the country based on the socio-economic status of the region and the localized nature of the knowledge. While accuracy was high for some subcultures, model performance was worse for regions with lower socio-economic status, and performance declined as knowledge became more culturally localized. My second study, an audit of caste and religious identity representation in LLM outputs, reveals that model defaults overwhelmingly favor socially dominant groups regardless of their numerical representation, and that light prompt-based interventions prove inconsistent in mitigating these biases. These two studies demonstrate that despite user feedback requesting corrections and diverse outputs, LLMs consistently favor information from dominant groups and globally popular cultural elements, and they do not adapt their generation to marginalized or locally specific perspectives. This reveals that while information about subcultures exists within the models, it doesn't surface readily, highlighting that the issue extends beyond missing data and includes problems with algorithmic generation. Finally, my third study shows that even on platforms (such as Snapchat) with uniform affordances across all users, cultural values drive significant differences in user behavior, such as friendship network structures and content consumption patterns (Stories). Specifically, compared to collectivism, individualism and high relational mobility lead to the formation of larger and less egocentric friendship networks and negatively moderate the effect of tie strength on content engagement. Collectively, these studies reveal that current algorithmic technologies inadequately serve diverse cultural contexts, demonstrating the need for a conscious effort to understand and incorporate cultural diversity into the data used for training and the design of algorithms that operate on this data to surface culturally diverse information. Finally, through its examination of cultural biases in LLM and cross-cultural user behaviors on social media platforms, this dissertation contributes to the growing field of culturally aware computing by advancing our understanding of LLM cultural alignment and bias, revealing patterns in cross-cultural digital behavior, and providing suggestions for frameworks that allow for developing more inclusive computational systems.
■590 ▼aSchool code: 0127.
■650 4▼aComputer science
■650 4▼aInformation science
■650 4▼aInformation technology
■653 ▼aResponsible AI
■653 ▼aNatural language processing
■653 ▼aCross-cultural AI
■653 ▼aHuman-computer interaction
■653 ▼aFairness and ethics in AI
■690 ▼a0723
■690 ▼a0984
■690 ▼a0489
■690 ▼a0800
■71020▼aUniversity of Michigan▼bInformation.
■7730 ▼tDissertations Abstracts International▼g87-03B.
■790 ▼a0127
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17359848▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


