Convolutional neural networks (CNNs) constitute the state of the art in automated tasks related to image analysis. The mechanism at the root of CNNs is the application of so–called convolutional filters: objects which play the ambiguous role of a “deforming sieve” that can be interpreted through the dualism between “philters” and “filters”. This article aims to investigate convolutional filters as the key ingredients in the “semiosic” process which occurs in CNNs. To this end I will detail how CNNs “see” their Umwelt that is their perceptual world, showing how figurativity emerges in their “mental representations”. I will focus on the phenomenon of “algorithmic pareidolia”, a fascinating example of the deformation a CNN forces on states of the world when it tries to recognise and classify them.