This paper proposes a new method to expose AI-generated fake face images or videos (Deep Fakes) by exploiting inconsistencies in head poses. Deep Fakes are created by splicing a synthesized face region into an original image, which introduces errors in facial landmarks. These errors can be revealed when 3D head poses are estimated from face images. The authors conducted experiments to demonstrate this phenomenon and developed a classification method based on this cue. Using features derived from this cue, an SVM classifier was evaluated using real face images and Deep Fakes. The Deep Fake production pipeline involves face detection, alignment, and synthesis. The synthesized face is then transformed back to match the original face. The 3D head pose estimation process involves calculating the rotation and translation of the world coordinates to the camera coordinates. The authors then compared head poses estimated using all facial landmarks and those estimated using only the central region. The results showed that the head poses for real images are close, while those for Deep Fakes are significantly different. The authors conducted experiments to confirm their hypothesis. They compared the cosine distance between head orientation vectors estimated using different landmarks. The results showed that the cosine distances for real images are significantly smaller than those for Deep Fakes. The authors then trained SVM classifiers based on the differences between head poses estimated using the full set of facial landmarks and those in the central face regions. The performance of the SVM classifier was evaluated on two datasets: UADFV and DARPA GAN. The results showed that the SVM classifier achieved an AUROC of 0.89 on the UADFV dataset and 0.843 on the DARPA GAN dataset. The authors also performed an ablation study to compare the performance of different types of features used in the SVM classifier. The results showed that including translation vectors improved the AUROC. The method is effective in detecting Deep Fakes by exploiting inconsistencies in head poses.This paper proposes a new method to expose AI-generated fake face images or videos (Deep Fakes) by exploiting inconsistencies in head poses. Deep Fakes are created by splicing a synthesized face region into an original image, which introduces errors in facial landmarks. These errors can be revealed when 3D head poses are estimated from face images. The authors conducted experiments to demonstrate this phenomenon and developed a classification method based on this cue. Using features derived from this cue, an SVM classifier was evaluated using real face images and Deep Fakes. The Deep Fake production pipeline involves face detection, alignment, and synthesis. The synthesized face is then transformed back to match the original face. The 3D head pose estimation process involves calculating the rotation and translation of the world coordinates to the camera coordinates. The authors then compared head poses estimated using all facial landmarks and those estimated using only the central region. The results showed that the head poses for real images are close, while those for Deep Fakes are significantly different. The authors conducted experiments to confirm their hypothesis. They compared the cosine distance between head orientation vectors estimated using different landmarks. The results showed that the cosine distances for real images are significantly smaller than those for Deep Fakes. The authors then trained SVM classifiers based on the differences between head poses estimated using the full set of facial landmarks and those in the central face regions. The performance of the SVM classifier was evaluated on two datasets: UADFV and DARPA GAN. The results showed that the SVM classifier achieved an AUROC of 0.89 on the UADFV dataset and 0.843 on the DARPA GAN dataset. The authors also performed an ablation study to compare the performance of different types of features used in the SVM classifier. The results showed that including translation vectors improved the AUROC. The method is effective in detecting Deep Fakes by exploiting inconsistencies in head poses.