COMPUTATIONAL ORIENTALISM: A COMPARATIVE FRAMEWORK FOR ASSESSING CULTURAL BIAS IN LARGE LANGUAGE MODELS

Abstract

As large language models increasingly mediate access to information about the Global South, understanding whether these systems reproduce colonial discourse patterns has become critical for both AI ethics and post-colonial studies. This thesis investigates whether the established 2025 benchmarks GPT-4 (OpenAI, USA) and Falcon3-7B-Instruct (Technology Innovation Institute, UAE) exhibit Orientalist discourse when generating content about the Middle East, testing whether developer geography shapes representational patterns or whether training data composition proves determinative. Drawing on Said's Orientalism alongside contemporary post-colonial scholarship, this research proposes an original framework, the Middle East Cultural Sensitivity Score (MECSS), independently conceptualized by operationalizing seven dimensions of Orientalist discourse—Homogenization, Agency Gap, Epistemic Center, Intelligibility Asymmetry, Temporal Asymmetry, Exoticization, and Legitimacy & Authority—into a systematic measurement framework. Analysis of 140 conversations per model (280 total; 1,120 conversational exchanges across both models) across seven theoretically grounded prompt categories reveals that both models reproduce Orientalist patterns systematically through structural positioning rather than explicit stereotyping. GPT-4 demonstrates moderate Orientalism (mean MECSS: 1.65), with Epistemic Center scoring highest (2.49), indicating Western analytical frameworks function as unmarked universals. Falcon3-7B-Instruct exhibits significantly higher Orientalism (mean MECSS: 2.16, +31%) despite its institutional development in Abu Dhabi and inclusion of Arabic content within its predominantly global, English-dominant training corpora, with particularly pronounced increases in Agency Gap (+0.96) and Homogenization (+0.67). This counterintuitive finding—that regional development amplifies rather than mitigates Orientalist patterns—supports the hypothesis that the statistical weight of general-purpose, global training data composition dominates over developer geography in shaping representational configurations. Theoretically, the research demonstrates Said's framework applies systematically to AI outputs, with Orientalism manifesting through epistemic structures embedded in training corpora rather than surface-level language. Practically, findings indicate that addressing AI bias requires fundamental transformation of what knowledge and frameworks structure training data, not merely expanding language coverage or relocating institutions.

Date of Award	2026
Original language	American English
Awarding Institution	HBKU College of Humanities and Social Science