MOSCOW, March 29 An image created by a neural network can be recognized by unnatural light or a distorted shadow, as well as by the irregular shape of hands, said the director for artificial intelligence and digital products of a mobile operator «Beeline» Konstantin Romanov.
«Recently, there is a lot of content that is generated by neural networks. It can be in the form of images or text … This content has recently been very difficult to distinguish from human-created content, but, as they say, «the devil is in the details.» As for images , while the technology of neural networks does not perfectly form a picture, and you can notice, for example, unnatural light or shadows that objects in these images cast, or just some random objects,» Romanov said.
As an example, the expert cites a recent analysis of a photograph of the Pope by Business Insider. «The wrong shadow, the distorted shape of the hand, and the extremely bulky down jacket were the points that made the photo 'viral,'» he said.
According to the expert, in the case of a text, it is possible to determine the material written by the neural network by inconsistent sentences and inconsistencies. “That is, if after a few sentences a description appears in the text of something that was not previously discussed, then this text was probably generated by artificial intelligence,” Romanov explains.
At the same time, as the expert notes, neural networks are initially trained in such a way as not to be distinguishable from a person.
«While there are details in the texts and images that clearly indicate the artificial origin of the content, but neural networks are improving and it will soon be more difficult to distinguish their creations from a person. That is why algorithms that are able to distinguish content generated by a neural network are becoming more and more popular today, from man-made», concludes Romanov.