The dissemination of fake news has become a significant concern in the current society. This problem is evident on social media platforms, where the spread of misinformation has become a constant presence in the daily lives of many individuals. In this work, we investigate the performance of the GPT-3.5 model in classifying fake and real news, considering 200 newspaper articles and two strategies for question formulation. Our results reveal that using a well-formulated question is crucial to obtain more precise responses. In particular, we observed an improvement of 21.1% in the F1-Score metric by directing the question to focus on the characteristics of a fake text.