In order to survive, animals constantly face decisions between competing interpretations for noisy and sparse sensory data. In statistics, multiple formal frameworks exist to tackle this problem, known as Model Selection (MS). However, despite the abundance of theoretical methods, the way in which MS is performed by the brain is poorly understood. Previous work showed that humans have a bias towards simpler models or interpretations of sensory data [1,2,3]. Human model selection strategies have been quantitatively described using a linear combination of the models Maximum Log-Likelihood ($L$) and terms penalizing their complexity [1,4], a method that can approximate many theoretical MS criteria (Fig b). In particular, humans showed a strong sensitivity to model Dimensionality ($D$): models with no degrees of freedom ($D=0$) were preferred to models with $D=1$. However, the functional shape of this bias, dependent on $D$ and possibly on the number of available datapoints ($N$), is still unknown. To address this question, we investigate how naïve humans perform MS when guessing the number of hidden sources of a noisy dataset. By asking subjects to decide between multiple models with different dimensionality (Fig a), we can measure how the strength of the bias changes as a function of $D$. Similarly, by showing trials with a varying number of datapoints ($N$), we can investigate its dependence on the amount of available data. Preliminary results show that on average the bias does not depend on $N$, as prescribed by MS criteria like the Akaike Information Criterion (AIC), and opposed to criteria like the Bayesian Information Criterion (BIC). However, the penalty imposed by subjects on models with a larger number of degrees of freedom appears to be overall sublinear in $D$, as opposed to the linear trend prescribed by both AIC and BIC (Fig c). This trend hints at a compressive Weber-like perception of Dimensionality. Another striking difference from classical MS frameworks was found in the way subjects assess the goodness of fit of models: instead of the extensive $L$ prescribed by most criteria, subjects behavior is better described by $L/N$ (Fig d), i.e. the cross-entropy between the empirical distribution of available data and the best fitting distribution of the model. Overall, our results show significant departures of human strategies from normative theories of MS, pointing towards an important role of cognitive resource limitations during intuitive MS.