Active sensing is a process using which animals improve their environmental perception and exploration at the cost of extra energy. Despite the ubiquitous nature of this phenomenon across animal taxa, the mechanisms initiating active sensing movements remain unknown. Existing models in the literature model active sensing either as an open-loop noise source or a stochastic resonance generator. In contrast, our previous work (Karagoz et al., 2024) model active sensing as a closed-loop process, where animals seek to improve its state-estimation performance by conducting additional active sensing movements. Despite strong correlations of these work with animal data, the validity of these models has only been tested in computer environments using offline data analysis. Here, we conduct animal-in-the-loop experiments by closing a virtual feedback loop around the animal using different models of active sensing. Specifically, we use a custom-built setup to investigate how Eigenmannia virescens, a species of weakly electric fish, reacts to perturbations of different active sensing generator models during their refuge tracking behavior. We manipulated (amplified or attenuated) the active sensing movements for N=3 individual fish using different active sensing generator models and observed how Eigenmannia adapts their tracking and active sensing movements. Our results show that the active sensing is a closed-loop process, where animals generate purposeful perturbations to their motor commands to improve their task-dependent state-estimation performance. This research is supported by TUBITAK (120E198).