Exploring the origins of animal and human behavior within cortical circuits remains a puzzle awaiting resolution. It is established that when animals and humans compare two stimuli (d1 and d2) separated by a delay interval (Fig. 1a), their perception is influenced by a perceptual bias, characterized by a contraction in the perception of the first stimulus towards its distribution’s mean (highlighted in red in Fig. 1a). To shed light on this effect, recurrent neural networks (RNNs) of spiking neurons were trained, and their behavior and firing activity were examined in comparison to those observed in humans and animals. The effect of contraction bias increases the difficulty for some classes (each pair (d1, d2)) and favors others. Performance of trained spiking RNNs (Fig. 1b) exhibited the contraction bias. Their behavior closely resembled that observed in humans and animals. Aligning with experimental findings, the networks' behavior was explicable within a Bayesian framework and conformed to Bayesian predictions, displaying a larger contraction bias with increasing noise [1] and longer delay intervals [2]. Furthermore, they replicated results from a psychophysics experiment where participants underwent reward protocols designed to either enhance or suppress the bias [3]. Analyses of the networks’ population activity showed that diverse geometric properties of trajectories (Fig. 1c) in neural state space encoded distorted representations of the first stimulus. Interestingly, these representations conveyed its Bayesian estimate [4], linking neural activity to behavior (Fig. 1d). To check these predictions, prefrontal cortex neural recordings taken in monkeys trained to compare the frequencies of two vibrotactile vibrations were analyzed [5], revealing similar Bayesian transformations on the first applied frequency. In summary, this study indicates that spiking RNNs align with experimental findings and accurately describe the relationship between population activity and behavior. These results suggest that a similar strategy could be employed both by the brain and by trained RNNs and in different cognitive tasks, generating biases through Bayesian computations. This underscores the potential of biologically realistic networks (i.e., spiking RNNs) to successfully replicate and elucidate the intricate mechanisms underlying human and animal behavior.