The pervasive presence of generative artificial intelligence in contemporary society has raised fundamental concerns about its alleged neutrality. This thesis examines the phenomenon of cultural bias within Large Language Models (LLMs) and Text-to-Image (TTI) models, with the aim of outlining the state of the art, taking into account the main manifestations of the phenomenon, evaluation methodologies and mitigation strategies. The literature highlights that, although there is evidence of a recurring alignment with values predominantly associated with the West and ‘WEIRD’ contexts, the phenomenon is more complex than a simple replication of a dominant culture. For example, LLMs can develop unique value systems that are far from real human cultures, and responses relating to cultural values can vary depending on a multitude of factors, including language, data, and training and development procedures. In TTI models, bias manifests itself through various representational distortions, such as exoticism, stereotyping, cultural misappropriation or patchwork. Furthermore, there is a lack of standardised methodologies and effective strategies, alongside critical trade-offs: reducing Western bias may conflict with respect for human rights in LLMs, whilst mitigating it may lead to overcorrections and historical inaccuracies in TTI models. Finally, possible directions for future research are outlined, such as the development of standardised evaluation metrics and new datasets that are more representative of cultural diversity, as well as the adoption of multidisciplinary approaches to help foster, within the field of AI, a digital landscape that is equitable, inclusive and respectful of the complexity of human cultures.
La presenza pervasiva dell’intelligenza artificiale generativa nella società contemporanea solleva criticità fondamentali circa la sua presunta neutralità. Questa tesi esamina il fenomeno del bias culturale all’interno dei Large Language Models (LLM) e dei modelli Text-to-Image (TTI), con l’obiettivo di delineare uno stato dell’arte, considerando le principali manifestazioni del fenomeno, metodologie di valutazione e strategie di mitigazione. La letteratura evidenzia che, sebbene sia documentato un ricorrente allineamento a valori associati prevalentemente all’Occidente e a contesti “WEIRD”, il fenomeno è più complesso di una semplice riproduzione di una cultura dominante. Ad esempio, gli LLM possono sviluppare configurazioni valoriali uniche e distanti dalle reali culture umane e le risposte relative ai valori culturali possono variare in funzione di una molteplicità di fattori, tra cui lingua, dati e procedure di addestramento e di sviluppo. Nei modelli TTI, il bias si manifesta invece attraverso varie distorsioni rappresentative, come l'esotismo, la stereotipizzazione, l'appropriazione indebita o il patchwork. Emerge, inoltre, l'assenza di metodologie standardizzate e strategie efficaci, insieme a trade-off critici: la riduzione del bias occidentale può entrare in tensione con il rispetto dei diritti umani negli LLM e la sua mitigazione può indurre ipercorrezioni e inesattezze storiche nei modelli TTI. Infine, sono presentate possibili direzioni per la ricerca futura, come lo sviluppo di metriche di valutazione standardizzate e nuovi dataset più rappresentativi della diversità culturale, oltre all’adozione di approcci multidisciplinari per contribuire, nell’ambito dell’IA, a un panorama digitale equo, inclusivo e rispettoso della complessità delle culture umane.
Bias culturale nell’IA generativa: stato dell’arte e prospettive future
MALVEZZI, FRANCESCA
2025/2026
Abstract
The pervasive presence of generative artificial intelligence in contemporary society has raised fundamental concerns about its alleged neutrality. This thesis examines the phenomenon of cultural bias within Large Language Models (LLMs) and Text-to-Image (TTI) models, with the aim of outlining the state of the art, taking into account the main manifestations of the phenomenon, evaluation methodologies and mitigation strategies. The literature highlights that, although there is evidence of a recurring alignment with values predominantly associated with the West and ‘WEIRD’ contexts, the phenomenon is more complex than a simple replication of a dominant culture. For example, LLMs can develop unique value systems that are far from real human cultures, and responses relating to cultural values can vary depending on a multitude of factors, including language, data, and training and development procedures. In TTI models, bias manifests itself through various representational distortions, such as exoticism, stereotyping, cultural misappropriation or patchwork. Furthermore, there is a lack of standardised methodologies and effective strategies, alongside critical trade-offs: reducing Western bias may conflict with respect for human rights in LLMs, whilst mitigating it may lead to overcorrections and historical inaccuracies in TTI models. Finally, possible directions for future research are outlined, such as the development of standardised evaluation metrics and new datasets that are more representative of cultural diversity, as well as the adoption of multidisciplinary approaches to help foster, within the field of AI, a digital landscape that is equitable, inclusive and respectful of the complexity of human cultures.| File | Dimensione | Formato | |
|---|---|---|---|
|
Malvezzi_Francesca.pdf
Accesso riservato
Dimensione
1.08 MB
Formato
Adobe PDF
|
1.08 MB | Adobe PDF |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/113691