In today's digitally interconnected world, the unchecked spread of rumors and misinformation presents significant challenges to public discourse, trust, and decision-making. Rumors, often based on unsubstantiated information, speculation, or exaggerated accounts, can rapidly disseminate through social media and other communication channels, particularly in uncertain or anxious situations. False rumors not only damage reputation and trust but also have tangible economic consequences, affecting businesses and markets.This research aims to provide a robust method for detecting rumors on social media platforms, addressing urgent requirements for effective techniques and tools to mitigate their impact. Initially, data are collected from social media and pre-processed. Word embeddings are then used to extract features from the preprocessed data, providing rich, context-aware vector representations that capture essential semantic and syntactic nuances for accurate classification. The extracted features are subsequently classified using a Generative Pre-trained Transformer (GPT), which relies on a transformer architecture that employs a self-attention mechanism to weighthe significance of unusual terms in a sequence comparative to each other. Finally, hyperparameters are optimized using the Deer Hunting Optimization Algorithm (DHOA). For experimental analysis, three types of datasets are utilized. The proposed method achieves high accuracy rates of 99.5% for Twitter15, 99.30% for Twitter16, and 99.1% for the PHEME dataset.