Is AI biased against people like me?

The fear, stated fairly

AI systems learn by finding patterns in enormous amounts of data, mostly taken from the internet. That data is not neutral. It over-represents some languages, cultures and viewpoints and under-represents others. If an AI system is used to screen job applications, recommend medical treatment, or answer a child's questions, its hidden assumptions could quietly disadvantage people who don't fit the patterns it learned. For Arabic speakers, women, older people, migrants and people with disabilities, the worry is that AI could reinforce old unfairness at new speed and scale.

What the evidence says

Bias against Arab culture has been measured. In a study presented at ACL 2024, one of the leading conferences in language technology, researchers at Georgia Tech built a benchmark called CAMeL, with 628 prompts and more than 20,000 culturally relevant names, foods, places and other entities. They tested 16 models, including some trained specifically on Arabic. The models tended to favour Western entities even when prompted in Arabic about Arab contexts, and sometimes produced stereotypes. The authors point to training data drawn heavily from sources such as Wikipedia as a likely cause (Naous et al., ACL 2024). The paper's title, "Having Beer after Prayer?", captures the kind of cultural mismatch they found.

Researchers take it seriously. More than half of the AI researchers in the largest survey of the field said that worsening inequality deserved substantial or extreme concern (Grace et al., JAIR, 2025).

What is genuinely risky

  • Invisible decisions. Bias does most harm where people don't know an AI system was involved, such as in hiring, lending or eligibility checks.
  • Language gaps. Tools that work well in English may work worse in Arabic, and worse again in Gulf dialects.
  • False confidence. A biased answer delivered fluently can be more persuasive than a hesitant human one.

Why there is hope, and what is being done

Bias can be measured, which means it can be managed. The CAMeL benchmark is itself a tool. Developers can now test their models for cultural bias against Arab contexts and track improvement over time.

Models built with regional data and institutions. Qatar's Fanar was developed at Hamad Bin Khalifa University with contributions from the Ministry of Endowments, Qatar National Library, Al Jazeera and others (HBKU). Saudi Arabia's ALLaM and the UAE's Jais were built as Arabic-first models. Local data does not automatically remove bias, and the CAMeL study found bias even in some Arabic-trained models. But it gives builders better material to work with, and it puts people who understand the culture in the room.

Principles and standards in the Gulf.

  • The UAE Charter for the Development and Use of AI (2024) names avoiding algorithmic bias among its 12 principles (UAE Government portal).
  • Saudi Arabia's data and AI authority, SDAIA, published AI Ethics Principles in 2023. These are voluntary guidance rather than binding law, and include fairness among their principles (secondary source: to be checked against SDAIA's published text).
  • Bahrain adopted the GCC Guiding Manual on the Ethics of AI Use in July 2025 alongside its government AI policy.
  • The Responsible AI Future Foundation, set up in Abu Dhabi in 2025 by G42, Microsoft and MBZUAI, aims to promote responsible-AI standards across the Middle East and Global South (Microsoft).

Human oversight still matters. Most responsible-AI frameworks, including the UAE Charter, call for human oversight of consequential decisions. That is the safety net when bias slips through.

What you can do

  1. Notice when answers don't fit. If an AI tool misunderstands your culture, language or situation, that is useful information. Report it where you can.
  2. Ask whether AI was used in decisions about you, such as a job application or a loan, and whether a person reviewed the outcome.
  3. Use Arabic-first tools where they suit you, and compare their answers with other tools.
  4. If you build or buy AI, test it on the people who will actually use it, in their languages and contexts.