Detecting fake news is crucial in mitigating the spread of misinformation on social platforms. However, this task remains challenging, particularly due to the prevalence of short, sparse, and noisy content in social media. This paper introduces GCN-BERT, a novel hybrid framework that integrates Graph Convolutional Networks (GCNs) with Bidirectional Encoder Representations from Transformers (BERT) to jointly model document-level structural relationships and contextual semantics. The proposed model constructs a graph of documents using TF-IDF similarity and learns structural embeddings via GCN, while BERT captures deep semantic representations. These features are fused and passed through a multilayer perceptron for classification. Experiments on a benchmark dataset show that GCN-BERT significantly outperforms the TF-IDF-SVM baseline, achieving an F1-score of 0.86 versus 0.70. The results highlight the effectiveness of integrating structural and contextual representations for robust fake news detection in short-text settings.