Abstract

Excerpted From: Bakht Munir, Islamophobic Artificial Intelligence in the U.S.: A Critical Analysis of Religious Bias in Datasets, 118 Law Library Journal 217 (2026) (109 Footnotes) (Full Document).

Artificial intelligence (AI) refers to the machine’s capability to execute tasks typically linked with human intelligence. Even though the term AI was initially used in 1956, there is no comprehensive definition of AI due to its multifunctionality. It may be defined as a non-natural entity’s capability of choice-making through an evaluative process. AI is increasingly integrating into various fields, revolutionizing every aspect of society. However, the transformation of AI confronts certain ethical concerns like bias.

A significant subset of AI is machine learning (ML), which allows learning without being directly programmed, historically traced back to the AI drive of the 1950s, emphasizing prediction and optimization. Deep learning is a kind of AI that employs artificial neural networks (ANNs) to train machines to process data in a way inspired by the human brain. It automates the tasks that traditionally require human intelligence. It is taught to generate new data that mimics the data they were trained on, establishing a correlation between the trained data and the prospective data. Hence, the outputs of generative AI largely depends on the underlying training data, and any bias in the models--whether racial, gender, or religious--is proportional to the biases inherent in that data.

Before considering Islamophobic AI, it is imperative to conceptualize religious bias and Islamophobia: religious bias refers to unequal treatment and prejudice for or against an individual or group grounded on their religious views or practices. It predisposes thoughts toward a particular religious belief and may manifest in various forms, including discrimination, stereotyping (generalized negative assumptions based on religion), harassment, and deliberate exclusion of individuals from activities or opportunities owing to their creed.

The rise in Islamophobia triggered discriminatory surveillance of Muslims in the Western world and created biases, substantiating state action against Muslim communities. Islamophobia is a negative emotion and illogical fear toward Islam and its followers. It is a prejudice, dislike, or terror directed toward Islam and may exist in the form of discrimination, negative stereotypes, hostility, or violence, mostly stemming from misapprehensions and fabricated information. In the context of AI, Islamophobic bias occurs when AI models exhibit negative outcomes due to prejudiced datasets or when the design of the AI model itself is influenced. Given the above conceptions, we can endeavor to define Islamophobic AI in the context of generative AI:

Islamophobic AI refers to the phenomenon where AI systems exacerbate religious bias against Islam and its followers in their decision-making process, perpetuating unfair treatment toward Muslims by generating negative stereotypes associating them with violence, or producing discriminatory content resulting in their inequitable treatment.

Studies have proven that AI models such as ChatGPT often link Muslims with violence more frequently than other religious groups, which may lead to real-world complications. Islam is the second-largest religion after Christianity, accounting for 25 percent of the global population. Hence, Islamophobic AI can imbalance global harmony and could be a significant threat to embracing AI worldwide. The proliferation of AI harbors Islamophobic biases, which can propagate detrimental stereotypes and prejudice against Muslims. Religious biases commonly stem from the data on which these models are trained. LLMs are expected to comprehend and imitate biases where the data on which they are trained encompasses biases or stereotypes. This article aims to investigate religious prejudices, particularly Islamophobic bias perpetuated in AI models, their impacts, and various debiasing techniques.

[ . . . ]

Islamophobic AI reflects the religious biases present in the datasets on which LLMs are trained, making generative AI inherently prone to bias.
Negative stereotypes and prejudices regarding a particular community based on its attributes like race, religion, or gender replicate biased generative content in AI models. A dataset motivated by religious bias in support of or against a particular religion and its followers will perpetuate unfair outcomes, which could lead to Islamophobic AI. It is challenging to defuse all kinds of biases from historical datasets;
however, different debiasing techniques may help mitigate biases. Most of the existing scholarship addresses negative stereotypes and undesirable Islamophobic content. A more challenging phenomenon emerges when the datasets are encrypted with intrinsic religious bias, although the datasets do not contain any explicit discriminatory or biased features. In these circumstances, the employment of scientifically established mitigating tools could not substantially address the proliferation of religious bias in the AI models’ outcomes. If the datasets on which the models are being trained on contain religious biases, AI will likely provide a prejudiced response.

To counter the religious bias in datasets, technological advancements coupled with machine learning models could significantly contribute to fair and unbiased AI operations by avoiding sensitive features in decision-making, overseeing and fine-tuning the results to prevent biases, confirming persistent performance through various groups, testing the outcomes in a hypothetical environment, and considering the collective effects of several features at the testing and production levels. These techniques constitute the foundation of algorithmic fairness, contributing to a more robust, equitable, and just AI system.

However, where a dataset is fundamentally biased, its generated superstructures cannot be fully debiased even with the optimum efforts of the designer. The AI systems are subject to training on large datasets across the globe. The designers cannot fully anticipate every expressed and implied religious bias and their complex nature to neutralize the data, accordingly, simultaneously maintaining its balance with valid criticism and the right to speech. Religious prejudices, in the given scenario of Islamophobic AI, are proportionate to society at large and can be debiased when society is optimized, free from negative stereotypes and religious biases. The intersection of AI in the legal field is in its infancy and rapidly evolving. Hence, an AI system designed by a diverse team with high-quality representative data is advisable where the end user is actively engaged to oversee and contribute to the model’s performance at the production level.

Postdoctoral Fellow, University of Kansas School of Law