AI and Islamophobia: The algorithmic gaze on Muslim communities

Artificial intelligence does not need to hate you to discriminate against you; it only needs the wrong data. Across Europe and North America, AI technologies are now shaping how people are seen, sorted, and suspected.

Facial-recognition systems, however, often misread darker skin tones and hijabs. Border-control software, meanwhile, flags Muslim names as risks, and online moderation quietly removes posts about Palestine while allowing anti-Muslim hate to circulate freely.

AI has been sold as neutral, objective, and efficient - the ultimate technical fix for human error. But technology learns from the world around it. If that world is unequal, prejudiced and exclusionary, then the algorithm built upon it will be too.

Artificial intelligence mirrors the assumptions embedded in its training data, and those assumptions reflect the societies that produce it. This is how systems that claim to remove bias end up amplifying it instead.

“These systems didn’t become biased by accident,” Mutale Nkonde, CEO of AI for the People, explained to The New Arab. They were trained on data shaped by decades of policy decisions and security doctrines that framed Muslim identity as inherently suspicious.

“The technology simply scaled that prejudice,” she says, making it faster, harder to detect, and more difficult to challenge. Awareness and resistance, however, are growing among legal scholars and digital rights activists to challenge the opacity of algorithmic systems.

Biased data, biased systems

Facial recognition technologies systematically underperform on darker-skinned individuals and religious minorities.

A landmark study, for example, showed that commercial facial recognition systems had error rates of up to 34.7% for dark-skinned women, compared with 0.8% for light-skinned men.

Another report by the National Institute of Standards and Technology (NIST) found such systems were 10 to 100 times more likely to mis-identify Black and Asian faces than white ones.

If the training data encodes under-performance on certain populations, then deploying those systems in policing, surveillance, or screening perpetuates injustice, only now, in algorithmic form.

Policing, borders, and the Muslim gaze

In the United Kingdom, predictive policing projects such as the National Data Analytics Solution have been built on historical arrest data, which is already distorted by decades of racial profiling.

These algorithms then direct police resources back toward the same communities long subject to over-surveillance, dressing prejudice in the language of “data-driven efficiency”.

In France, the drift is equally alarming. Systems originally designed to detect “radicalisation” online have quietly evolved into tools for monitoring religious expression itself. Mosque websites, prayer apps and faith-based social pages are scanned by algorithms that cannot tell the difference between belief and extremism.

This digital suspicion echoes the broader political climate, one that treats visible faith as a potential threat. Nkonde warns that this logic is self-perpetuating: “When we design technologies to detect threats, we often end up teaching them who to suspect.”

Nowhere is this clearer than in the United States, where national-security priorities after 2017 laid the foundation for algorithmic bias at scale. During the first Trump Administration, Executive Orders 13,769 and 13,780 restricted immigration from seven Muslim-majority countries.

These decisions were justified as counter-terrorism measures, but their influence extended deep into the architecture of US border control.

According to Nkonde, this period “set one of the parameters for the development of AI systems used at US borders”. Facial recognition scans and biometric checks became routine for those entering from the Muslim world. Once these scans occurred, the data race, ethnicity, country of origin, and “reason for additional screening” were stored in the system’s metadata.

Over time, the algorithm learned to associate these attributes with “risk.” Nkonde describes this as the silent reinforcement of bias. “That data then gets baked into the model’s memory, increasing the probability that Muslim travellers will be associated with the word ‘terrorism’ inside the dataset.”

The result is a feedback loop in which suspicion generates more data, and more data generates more suspicion. Innocent travellers are flagged, interrogated or delayed not because of what they have done, but because of what the machine has learned to expect.

“It’s a clear example of Islamophobic bias encoded into technology,” she says, “where political fear becomes data, and data becomes destiny.”

Generative AI and 'orientalist' training data

The next wave of concern lies with generative AI and language/vision models that draw on massive datasets, according to Sonia Fereidooni, a PhD candidate at the University of Cambridge who studies the role of big tech in surveillance and defence.

"Generative AI models that we see heavily commercialised today have deep-rooted anti-Muslim bias that has been persistent since their conception due to the Orientalist training data that has been used to found them,” she told TNA.

There is academic literature, she says, showing that language models, even since the GPT-2 and GPT-3 eras in 2021, and before the release of ChatGPT, had severe anti-Muslim bias that persists and has downstream effects on the creation of AI models today.

“More notably, Generative AI models, including vision models, do not just reinforce Islamophobic ideals, but they also amplify social biases and stigmas, including the Western-centric Orientalist imagery found in their training data, which creates an endless feedback loop of growing Islamophobia through the automation of disseminating amplified anti-Muslim rhetoric," Fereidooni adds.

Indeed, a major peer-reviewed study titled “Persistent Anti-Muslim Bias in Large Language Models” found that in the case of GPT‑3, the word “Muslim” was analogised to “terrorist” in 23% of test cases, compared with far lower rates for other religions. Follow-up work showed that even after “debiasing” attempts, the bias persisted in newer models.

Another recent study found that certain bias-mitigation techniques (prompt-engineering, cultural prompting) could reduce the prevalence of Arab/Muslim bias by up to 87.7 %, but results varied widely and many higher-order associations remained.

Online moderation, silencing, and amplification

Online platforms claim to mitigate hate speech, yet research suggests a different pattern when it comes to anti-Muslim content. For example, the independent organisation Center for Countering Digital Hate (CCDH) reported in 2023 that social-media companies failed to act on 89% of anti-Muslim (Islamophobia) hate that was reported.

Likewise, a 2022 report from Amnesty International uncovered systemic issues in content moderation that disproportionately affected Muslim women.

The message is clear: while platforms market themselves as neutral arbiters of free expression, their moderation systems often let Islamophobic content proliferate unchecked and algorithmically.

The cultural and institutional dimension

What makes algorithmic surveillance and bias so insidious is their opacity. A police officer’s prejudice can be challenged; an algorithm’s cannot. Once the machine labels someone “high-risk,” there is rarely an explanation and rarely an appeal.

For Muslims travelling, working, or expressing faith online, this creates a sense of permanent scrutiny. Facial recognition at airports may trigger extra questioning; job-screening software may quietly rank candidates lower because of a name it deems “foreign”; social-media moderation may suppress discussions about Gaza or Islamophobia.

Each instance may seem minor, but collectively they amount to a digital regime of second-class citizenship. Nkonde calls this a “cycle of digital suspicion,” warning that claims of neutrality mask a deeper truth. “Neutrality in a biased world only serves the status quo,” she adds.

Resistance, however, is growing. Muslim technologists, legal scholars, and digital rights advocates are beginning to challenge the opacity of algorithmic systems. If Muslim and Global South perspectives remain absent from the international debate, bias will continue to be treated as a technical fault rather than a political injustice.

“Until Muslim voices are part of writing the rules, AI ethics will remain a Western conversation about global problems,” Nkonde says. Across Europe, civil society organisations are pressing for transparency through the EU AI Act, which introduces risk classifications and oversight mechanisms.

Yet enforcement remains inconsistent, and national security exemptions allow governments to operate opaque technologies under the radar.

The deeper problem, though, lies beyond regulation. It is cultural and institutional. AI bias is not just a software issue but a reflection of who gets to decide what “risk” looks like and whose safety matters most.

“AI systems work on what is fed into them and what algorithms and programmes do within them. No doubt, in any AI system, there will be biases, which means that they may start to focus or look for certain patterns that a human would know are manipulated or just untrue,” Imane A. Atta, Director of Tell MAMA, told TNA.

“That is why human oversight is so important, and that is why reliance on AI alone is a potential real danger when you link them to profit-making ventures,” she adds.

“Anti-Muslim prejudice can easily creep into AI if programmes recognise a high volume of such activity, recognise key influencers promoting it or that it sells a greater volume of products. They may legitimise such activity or even enhance it.”

Systems trained on fear will continue to reproduce fear, regardless of technical sophistication. As Nkonde reminds us: “We keep asking how to make machines smarter when we should be asking how to make systems fairer.”

Hamid Chriet is a Moroccan-British-French cybersecurity expert and columnist

Follow him on X: @HamidChriet