This study presents a causal approach for evalu- ating automated generation of reading passages for early readers. It consists of estimating the average treatment effect of a candidate model, then conditional average treatment effects with causal forests. We apply it to evaluate a fine- tuned model for a digital literacy platform