Generalization as the use of past experience to solve new problems is realized through the formation of internal representations of the external world, which can also be called reflection in a broad sense. A suitable model object for studying reflection is the simplest recurrent neural networks (RNN). We assume that agents capable of reflection can demonstrate a universal reflection skill, which means that after training on one reflection task, the agent solves another, previously unknown (test) task better than one who has not been trained at all. If such an effect is observed for different primary and test tasks, then we can confirm the presence of reflexive equifinality. In this paper, we test the presence of reflexive equifinality by considering two reflection tasks: reflexive games (even-odd, rock-paper-scissors) and responding to fixed time series of stimuli according the mentioned game rules. The formation of a universal reflexive skill was not observed, since the trained RNN coped with new tasks as well or worse than the untrained ones. However, responding to fixed time series allowed the RNN to maintain its ability to adapt to new conditions better than the reflexive game, which was also shown from the point of view of the structural characteristics of the RNN. Therefore, in order to form internal representations, subjects need regular changes in environmental conditions. The obtained results help us to systematize our understanding of reflection and simplify the choice of environmental conditions, i.e. tasks, for further research of this cognitive phenomenon.