That is not expected behavior unless... do you have primers/linker/non-bio-nts in your reads? If so, you need to remove them (whether by trimming off the primer with trim-left or with a dedicated tool).
DADA2 (and any method that infers exact sequences) will interpret the ambiguous nucleotides that are in most primers as biological variation, which can artefactually multiply the number of features you get out at the end. For normal data, w/o primer sequence, the number of features should change relatively little when chopping off 10 nts.