One of the tasks assigned by man to artificial intelligence is the recognition and generation of images. In the process of solving such problems, among obvious and predicted errors, specific ones appear — distinctive only to the results of AI. The most popular image search services do not have a single database of source images. AI in these services is trained according to the reaction of users, their actions affect the recognition results and subsequent work.

In the post-Soviet space, the Yandex image search service is the most common. Yandex, like many similar systems, is constantly learning through the reaction of people to the result that it produces. In the process of self-education, two polar side effects occur. Some users of the service look for visual information about the tragedies, teaching AI to connect simple images with drama and developing its phantasmatic abilities. The other part searches for simple things, teaching AI to ignore image complexity, for example: compositional and contextual features.

Trying to recognize the image, the AI ​​divides it into the smallest objects, sequentially simplifying the task and forming new images. This process resembles synesthetic vision. The variety of possible recognition results is gradually reduced, but often at the end of ​​search AI is alone with the problem of choosing between a bad and a very bad option. «Not Found» — is an acknowledgement of failure.

Projects