AI-generated content is becoming more prevalent which poses an important threat to the authenticity of visual media. The paper compares deep learning and classical machine learning methods to identify deepfake images. We investigate transfer learning with a fine-tuned EfficientNetB0 model along with logistic regression. Both approaches are validated on a publicly available dataset of real and AI-generated images. Results also indicate that EfficientNetB0 beats logistic regression with 92% vs. 56% accuracy for logistic regression. The results indicate that in this context deep learning is more important and also suggest the trade-offs between model complexity, interpretability, and performance. This is useful practical guidance into establishing rigorous deepfake detection mechanisms, which are indispensable in order to alleviate the risks of synthetic media.