During model training, the detector looks at a specific set of images which is assumed to cover a wide range of real and fake media. But as deepfake technology advances, it is difficult for the trained detector to generalize and perform well on all of these new methods instantly. In addition to this, there are many variations/transformations a media goes through in the real world such as passed through social media, resized, etc. This adds additional noise into the media, which makes the job of the detector harder and can lead to wrong conclusions.
It is worth noting that the Reality Defender Research team is constantly evaluating new methods of deepfaking and training/adding models to our platform, often before these methods make it to mainstream usage.