Prompt Engineering Techniques for Mitigating Cultural Bias Against Arabs and Muslims in Large Language Models: A Systematic Review

Bushra Asseri, Estabrag Abdelaziz, Areej Al-Wabil · 2025

Read Prompt Engineering Techniques for Mitigating Cultural Bias Against Arabs and Muslims in Large Language Models: A Systematic Review by Bushra Asseri, Estabrag Abdelaziz, Areej Al-Wabil (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2506-18199.txt.

Introduction

Bushra Asseri, Estabrag Abdelaziz, and Areej Al-Wabil review prompt methods to cut cultural bias in language models, in a 2025 paper. Bias against Arabs and Muslims can spread stereotypes and push groups to the margins. Prompt design is a low cost lever, but work aimed at these groups was understudied.

The review pools eight studies from 2021 to 2024 under standard review rules. Five prompt approaches stand out, from cultural framing to multi step pipelines. All show promise, though strength varies by bias type and some bias resists prompt fixes. Multi step pipelines cut bias the most in the set, while cultural prompting is easier for broad use.

Catalog record

CREATORBushra Asseri, Estabrag Abdelaziz, Areej Al-Wabil
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE60 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2506-18199

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Prompt Engineering Techniques for Mitigating Cultural Bias Against Arabs and Muslims in Large Language Models: A Systematic Review Bushra Asseri, Estabrag Abdelaziz, Areej Al-Wabil Abstract Large language models (LLMs) have demonstrated remarkable capabilities across various domains, yet concerns about cultural bias—particularly towards Arabs and Muslims—pose significant ethical challenges by perpetuating harmful stereotypes and marginalization. Despite growing recognition of bias in LLMs, prompt engineering strategies specifically addressing Arab and Muslim representation remain understudied. This mixed-methods systematic review examines such techniques, offering evidence-based guidance for researchers and practitioners. Following PRISMA guidelines and Kitchenham’s systematic review methodology, we analyzed 8 empirical studies published between 2021-2024 investigating bias mitigation strategies. Our findings reveal five primary prompt engineering approaches: cultural prompting, affective priming, self-debiasing techniques, structured multi-step pipelines, and parameter-optimized continuous prompts. Although all approaches show potential for reducing bias, effectiveness varied substantially across studies and bias types. Evidence suggests that certain bias types may be more resistant to prompt-based mitigation than others. Structured multi-step pipelines demonstrated the highest overall effectiveness—achieving up to 87.7% reduction in bias—though they require greater technical expertise. Cultural prompting offers broader accessibility with substantial effectiveness (71-81% improvement in cultural alignment). These results underscore the accessibility of prompt engineering for mitigating cultural bias without requiring access to model parameters. The limited number of studies identified highlights a significant research gap in this critical area. Future research should focus on developing culturally adaptive prompting techniques, creating Arab/Muslim-specific evaluation resources, and integrating prompt engineering with complementary debiasing methods to address deeper stereotypes while maintaining model utility. keywords Large Language Models, Prompt Engineering, Cultural Bias, Arabs, Muslims, AI Ethics 1 Introduction Large language models (LLMs) have substantially advanced natural language processing by generating human-like text, answering questions, and performing a wide array of language tasks. Yet, research consistently demonstrates that these models can perpetuate—and even amplify—societal biases embedded in their training data ( Bender et al. 2021 ; Blodgett et al. 2020 ) . Of particular concern is cultural bias against Arabs and Muslims, which has been documented across various LLMs and poses significant ethical challenges in real-world deployments ( Abid et al. 2021 ; Naous et al. 2024 ) . This bias often takes the form of associating these groups with terrorism, violence, or religious extremism, thus reinforcing damaging stereotypes that can fuel marginalization and discrimination. A key lens for understanding how LLMs inherit and reproduce biases is captured in Bender et al. 2021 influential framing of these models as “stochastic parrots.” This term underscores the idea that LLMs, no matter how sophisticated, essentially generate tokens based on probabilistic patterns gleaned from massive training corpora. LLMs do not possess genuine comprehension of meaning; rather, they recycle and remix the biases, stereotypes, and prejudices that are baked into their source data. Consequently, these models risk echoing or even exacerbating the cultural and religious biases present in online text and other data repositories. By focusing on statistical mimicry instead of semantic understanding, LLMs become conduits for embedding longstanding social hierarchies and inequities into automated systems, raising critical questions about ethics and responsibility in AI-driven technologies. To fully grasp why cultural bias against Arabs and Muslims emerges so strongly in LLMs, it is useful to consider Said 2019 concept of Orientalism. Said argues that Orientalism is not merely a set of negative stereotypes but rather a Western scholarly and cultural framework that has historically depicted Arab and Muslim societies as exotic, backward, or dangerous “others.” This process of “othering” served to justify and maintain colonial power structures, shaping cultural discourse in ways that outlasted the colonial era. Given that much of the data used to train LLMs originates from Western-centric sources, these Orientalist tropes can be inadvertently reinforced and reproduced by the models. Instead of neutral or objective representations of Arab or Muslim identities, LLM outputs may mirror the historical power imbalances and cultural misconceptions enshrined in the training corpus—thus perpetuating Orientalism in contemporary digital forms. Beyond the high-level insights provided by Bender et al. 2021 and Said 2019 , researchers have developed more granular frameworks for understanding the mechanics and ethics of bias in AI. Hovy and Prabhumoye 2021 propose a structured approach that identifies five sources of bias in natural language processing systems: data, the annotation process, input representations, models, and research design. While this framework underscores the multifaceted origins of bias, Prakash and Lee 2023 further extend this analysis through the notion of “layered bias,” highlighting how bias can be embedded not just at the dataset or annotation level, but also within the internal layers of transformer architectures themselves. Interventions must therefore reach beyond the surface (e.g., fixing a flawed training dataset) to grapple with how biases permeate deeper representational strata. From an ethical standpoint, Birhane 2021 calls for a shift from rational to relational thinking, emphasizing that technical fixes alone are insufficient to address systemic bias. This relational ethics perspective encourages examination of power structures, historical injustices, and lived experiences, resonating strongly with Said 2019 critique of how Western knowledge production has historically marginalized Arab and Muslim voices. Indeed, “stochastic parrots” can reflect Orientalist biases if the training data—and thus the very fabric of these models—repeatedly portrays Arabs and Muslims through a limited, negative lens. Among the different types of biases studied in LLMs, cultural and religious bias against Arabs and Muslims has received increasing scholarly attention ( Abid et al. 2021 ; Naous et al. 2024 ) . Such bias can manifest both explicitly (e.g., negative or stereotypical completions to prompts) and implicitly (e.g., subtle misrepresentations of cultural practices). The concept of “stereotype dimensions” ( Schuster et al. 2025 ) helps clarify how models encode and perpetuate stereotypes about Arabs and Muslims through contextual embeddings, while the “analogical bias theory” ( Manzini et al. 2019 ) illustrates how these stereotypes become ingrained via analogy-based relationships in embeddings. Critically, these studies also highlight the intersectionality of cultural, religious, and other identity factors, making mitigation efforts all the more nuanced. The persistence of these biases is especially troubling as LLMs become embedded in global applications such as virtual assistants, automated content moderation, and policy-making tools. Abid et al. 2021 further show that anti-Muslim bias in LLMs can fluctuate in tandem with geopolitical events and media coverage, increasing by up to 18% during periods of heightened attention to terrorism incidents. This underscores the need for agile mitigation strategies that can adapt to rapidly shifting social and political contexts—something mere dataset-level interventions may struggle to achieve. Given the diverse sources of bias ( Hovy and Prabhumoye 2021 ) and their layered nature ( Prakash and Lee 2023 ) , prompt engineering has emerged as a compelling approach for mitigating problematic outputs, particularly in scenarios where users lack direct access to model parameters. By carefully crafting or modifying input prompts, users can guide LLM behavior without retraining the model—a critical advantage for closed-source commercial systems ( Gallegos et al. 2024 ; Meade et al. 2022 ) . Moreover, prompt engineering offers both accessibility (allowing non-technical stakeholders to participate) and adaptability (facilitating quick responses to emerging biases). However, the effectiveness of prompt engineering must be contextualized within broader ethical and sociopolitical frameworks. Birhane 2021 relational ethics challenges overly technicist solutions, arguing that any mitigation technique—prompt engineering included—must

Read the complete text