The striking realism of AI-created pictures—usually prompt-based—has muddied the divide between real and fake images, with serious implications for disinformation and trust. To deal with this, we compiled a new, diverse image set consisting of both actual photographs and AI-created items. We then tried out well-known pretrained CNNs (ResNet, VGG, EfficientNet), initially utilizing their learned features off-the-shelf and then fine-tuning them to more specifically adapt to this task. What we discovered was evident: fine-tuning these models to the data not only made them better at picking up on imperceptible, AI-only artifacts but also enhanced detection performance as a whole. With meticulous analysis of precision, recall, and F1-score, our research highlights how fine-tuning closes the gap between vision models that are general purpose and the subtle problem of fake image detection. By so doing, we are offering a workable benchmark and an example to further research in protecting authenticity in all forms of digital media.