Russian researchers from Samsung report that they have developed a new AI with a proudly swollen chest that can create deceptively real-looking video fakes from just a few photographs.
Deepfakes are on the one hand technically fascinating, on the other hand a grandiose possibility to ruin existences and exercise power. These are videos generated by artificial intelligence in which people perform and perform things without these people ever having done it.
This is made possible by learning algorithms. Roughly speaking, the algorithm learns to recreate motion sequences in video form and to combine these motion sequences with photos with which it is fed. The most groundbreaking thing about Samsung’s new artificial intelligence is that it manages with very few photos to make a certain person appear in such a deepfake video. For a great result 32 photos of the person are necessary. A single photo is sufficient for an acceptable result.
In plain language, this means: In case of doubt, a single Facebook profile picture is sufficient to let person X do any things in any video.
Deceptively real-looking deepfakes are associated with the problem of recognizing them as such. In the political context, this could be highly dangerous – if one can no longer distinguish politician X, who actually says something, from an AI-generated technical representation of politician X, to whom something is put in his mouth, there are innumerable possibilities for manipulating public opinion and influencing events in general. Of course, all this also works on other levels.
Which developments Samsung’s new super technology paves the way for, we cannot foresee. However, it is to be feared that the improved deepfake technology will also serve unfair purposes. A foretaste of this can be found, for example, in last year’s Deepfake, in which a fake Barack Obama attacks the current US President Donald Trump – optically it is hardly recognizable that it is a fake video.