Abstract Urdu fake news detection (FND) remains an under-resourced problem despite Urdu being spoken by over 231 million people worldwide. While prior work has demonstrated strong in-domain performance on individual Urdu datasets, whether models trained on one corpus generalise to another has received little systematic attention. This paper presents the first cross-dataset generalisation study for Urdu FND, using two publicly available balanced datasets: the Ax-to-Grind Urdu corpus (10,083 articles, 15 domains) and the Notri-Fact Urdu dataset (13,388 articles). We fine-tune xlm-roberta-base [1] under four experimental conditions - in-domain on each dataset, and two zero-shot cross-domain transfer directions - and compare against TF-IDF baselines using Logistic Regression (LR) and Support Vector Machines (SVM). Our experiments reveal a striking asymmetry: while B → A transfer achieves a macro F1 of 0.771, A → B transfer collapses to an F1 of 0.005, with the model predicting fake for 99.7% of all test articles. Through class-conditional length analysis and predicted label distribution inspection, we demonstrate that this collapse is attributable to a systematic length confound in the Ax-to-Grind dataset: fake articles average 117 words versus 35 words for real articles - a 3.4 × asymmetry that induces shortcut learning . Because Notri-Fact articles are uniformly long across both classes, the Ax-to-Grind-trained model has no valid length signal to apply at test time. These findings have direct implications for dataset construction standards and evaluation practices in low-resource NLP . We further provide a diagnostic methodology for identifying confound-driven model behaviour that is reusable across multilingual FND settings.