Does language development drive language change? A common account of language change attributes the regularity of certain patterns to children's learning biases. The present study examines these predictions for change-in-progress in the use of "must" in Toronto English. Historically, modal verbs like "must" start with root (deontic) meanings, eventually developing epistemic (probability) meanings in addition. Epistemic uses increase over successive generations, phasing out root uses (INCREMENTATION). The modal becomes unambiguously epistemic and eventually disappears from the language. Such cyclic changes are predictable and common across languages. To explore whether children contribute to incrementation and loss, we tested intuitions about "must" in preschoolers (n = 141) and adults (n = 29). In a picture-preference task (deontic vs. epistemic), children selected epistemic interpretations of ambiguous sentences (e.g., "Michelle must swim") at higher rates than adults. Two context-based preference tasks tested children's overall sensitivity to the presence of modals. We found sensitivity in deontic contexts. In epistemic contexts, where "must" is optional and functions like an evidential marker, we found little discrimination, and general avoidance of the modal. These results (epistemic overgeneration, "must"-avoidance) correspond to predictions of the incrementation hypothesis, suggesting children likely play an active role in language change, beyond well-known over-regularization processes.