In a groundbreaking advancement for the fight against digital misinformation, researchers have developed a deep learning model capable of detecting fake news in Urdu with an impressive 96% accuracy rate. This achievement marks a significant milestone for automated fact-checking systems in languages considered low-resource, such as Urdu, spoken by over 170 million individuals globally and ranked as the tenth most spoken language worldwide. Published in the prestigious journal Science Advances, this new study outlines the development and efficacy of an artificial intelligence (AI) system meticulously trained on a vast dataset of more than 14,000 Pakistani news articles, meticulously curated to encompass a wide spectrum of societal topics.
The unique challenge of building such a model stems primarily from the scarcity of comprehensive Urdu-language datasets, especially ones that adequately address sensitive and influential subjects such as politics and religion. Prior datasets, researchers found, suffered from severe limitations that excluded much of the potentially misleading content central to shaping public opinion and societal dynamics in Pakistan and its diaspora. By addressing these deficiencies head-on, the current research team has filled a critical gap, emphasizing content domains that have been overlooked due to cultural sensitivities.
The study’s lead author, Dr. Muhammad Zeeshan Babar from Heriot-Watt University in Edinburgh, Scotland, highlights a fundamental discrepancy in the global landscape of misinformation detection systems: most models are rooted in English language datasets. In contrast, Urdu’s paucity of large, diverse, and labeled datasets has historically stymied efforts to build similarly robust AI frameworks. This project’s approach not only rectifies this imbalance but also sets a precedent for future efforts aimed at languages that are currently underrepresented in the artificial intelligence arena.
To create the dataset, Dr. Babar and his collaborators systematically aggregated over 14,000 Urdu news articles authored between 2017 and 2023. These articles spanned 15 thematic categories, including politics, health, business, education, sports, science, crime, technology, and social issues, reflecting the multidimensional nature of news consumption among Urdu speakers. The dataset was carefully annotated, with 8,283 articles verified as factual and 5,895 identified as fabricated or misleading. This diverse compilation facilitated the AI’s ability to learn nuanced distinctions in language usage that differentiate authentic journalism from misinformation.
The AI model employs a sophisticated blend of state-of-the-art machine learning techniques, utilizing concatenated embeddings from BERT and GloVe. BERT (Bidirectional Encoder Representations from Transformers) provides a deep contextual understanding of the nuances in language, capturing the syntactic and semantic relationships within Urdu text. Meanwhile, GloVe (Global Vectors for Word Representation) contributes a comprehensive representation of word co-occurrences, enhancing the model’s grasp of semantic patterns. The synergy of these embeddings equips the model to detect subtle variations in vocabulary, phrasing, sentiment, and linguistic structure that often betray fabricated news stories.
In practical terms, this system can flag not only outright falsehoods but also misleading narratives and stories with partial truths, a critical feature given the complex nature of misinformation that often blends fact and fiction. This multi-dimensional detection capability addresses a significant shortcoming of prior models that focused solely on binary distinctions between true and false. The nuanced approach makes it a powerful tool for combatting viral misinformation that can influence public health decisions, electoral outcomes, and societal trust in government institutions.
Dr. Waseem Abbasi, head of computer science at the University of Lahore, praised the researchers’ commitment to transparency and collective progress by making the dataset open access. He emphasized the importance of community collaboration to iteratively enhance the model’s accuracy and adaptability to emerging misinformation trends. Despite achieving 96% accuracy—a considerable feat—he acknowledges that this still leaves room for error, which could have serious implications in domains like content moderation and legal matters.
Moreover, the team is keenly aware of the limitations posed by training algorithms on historical data, which may hinder their ability to recognize new types of misinformation, satire, or legitimate political dissent. These challenges underscore the importance of ongoing refinement and contextual understanding in AI systems to ensure they do not inadvertently censor or misclassify legitimate speech. For the millions of Urdu speakers navigating today’s complex media landscape, this AI represents a potentially transformative tool that can provide critical support in discerning trustworthy information.
The system’s potential extends beyond Urdu, as the research team has indicated that future projects will explore expanding this methodology to other languages that similarly lack substantial datasets for automated fake news detection. This vision promises to democratize the benefits of artificial intelligence in combating misinformation on a global scale, particularly in communities that have been underserved by current technological efforts dominated by English-centric resources.
Funded by Heriot-Watt University, this research exemplifies an intersection of advanced computational techniques and socially impactful objectives. It highlights the critical role of linguistic inclusivity in AI development and the profound societal implications of technological innovation geared toward preserving truth and trust in information.
The publication in Science Advances not only reflects the scientific rigor of the work but also ensures that the findings are accessible to a broad audience of researchers and practitioners interested in applied sciences, computer science, mass media, and communications. As misinformation continues to pose a grave threat worldwide, this breakthrough serves as a beacon of hope, demonstrating how technology can adapt to the linguistic and cultural complexities of diverse global populations.
Subject of Research: Fake News Detection in Urdu Using Deep Learning
Article Title: Verifying Urdu news authenticity using deep learning with concatenated BERT and GloVe embedding
News Publication Date: 5-Feb-2026
Web References: Science Advances Article
Keywords
Urdu language, fake news detection, misinformation, deep learning, BERT, GloVe embeddings, natural language processing, low-resource language, AI ethics, political misinformation, dataset development, media trust
