Standard model selection criteria, such as AIC and BIC, balance goodness-of-fit against complexity by contrasting an accumulated log-likelihood that scales linearly with sample size against a penalty term that grows only sublinearly, if at all. We investigated whether human intuitive inference mathematically adheres to this scaling principle when visually estimating the number of latent components in a Gaussian mixture model. By estimating subjects' implicit scaling functions via Gaussian Processes, we found that human evaluation of accumulated evidence sharply deviates from normative linear growth, exhibiting instead a markedly sub-linear trend. This indicates that intuitive human model selection doesn't rely on theory-like computations, but instead employs bounded cognitive heuristics that systematically underweight the impact of large sample sizes.