UNCOVERING DIGITAL DISCRIMINATION: ALGORITHMIIC BIAS AND DIVINE VALUE RESPONSES IN THE BIG DATA ERA

Abstract

AI systems now decide what content people see online, at scale. The engineers building them generally aim for fairness. The systems still discriminate against specific groups anyway. This study measures that bias in content moderation systems, focusing on religious, racial, and gender identity, then reads the results through Islamic ethics. We used a quantitative-descriptive approach and computational content analysis on the Jigsaw Unintended Bias in Toxicity Classification dataset 90,902 entries after balancing classes via proportional undersampling and measured bias with chi-square tests, independent t-tests, and fairness metrics including Demographic Parity Difference and Disparate Impact Ratio. The headline result: comments mentioning Muslim identity got approved 21.1 percentage points less often than comments mentioning Christian identity, 66.3% versus 87.4% (chi-square = 184.98, p < 0.001; Disparate Impact Ratio = 0.758, below the 0.8 line normally used to flag adverse impact). Three things in the training data drive this: annotation bias, historical context bias, and co-occurrence bias. We could check the third one directly the word "terrorist" shows up alongside Muslim-identity comments 13.3% of the time, against 2.5% for Christian-identity comments. We then read these results against four Islamic principles tabayyun (verification), 'adl (justice), musawah (equality), and amanah (accountability) within the Maqashid al-Shariah framework, which we define here as normative principles drawn from the Qur'an and Islamic jurisprudence governing just conduct, equality, and accountability. The paper's contribution is an Islamic AI Governance Framework that turns these principles into actual mechanisms: bias auditing, annotator diversification, algorithmic transparency, inclusive data governance.