Abstract
The explosive growth of social media has intensified the need for reliable toxicity-detection systems that keep online spaces safe and inclusive. Yet low-resource languages such as Kazakh remain under-served because limited data, agglutinative morphology and cultural nuance exacerbate algorithmic bias and privacy leakage. We introduce a fairnessand privacy-aware framework that combines named-entityrecognition (NER) anonymisation, adversarial debiasing and multi-task learning. First, the $\mathbf{1 0 0}$ k-comment Jigsaw Toxicity corpus is machine-translated into Kazakh and partially backtranslated for quality control. A multilingual DistilBERT encoder processes tokenised text, while an NER module masks PERSON, LOCATION and ORGANISATION entities to protect users. A gradient-reversal branch discourages the encoder from encoding sensitive attributes, and a lightweight reject-option layer adjusts borderline decisions in favour of under-represented groups. Comprehensive evaluation on the translated test set shows the proposed model reaches 92.45% accuracy, 75.90% precision, 72.50% recall and 74.15% F1-score. At the same time it lowers Equal-Opportunity Difference by 26%, detects 80% of sensitive entities and preserves 85% anonymisation consistency. The results demonstrate that strong predictive performance, fairness and privacy can be delivered simultaneously in a lowresource setting, offering a practical path toward ethical naturallanguage processing for the Kazakh digital ecosystem.