Working memory is necessary for a wide variety of cognitive abilities. Developmental work has shown that as working memory capacities increase, so does the ability to successfully perform other cognitive tasks, including language processing. The present work demonstrates the effects of working memory availability on children's language production. Whereas most of the previous research linking working memory to language development has been correlational, we experimentally varied the working memory load during concurrent language production in children 4-5 years old. Participants were asked to describe simple picture scenes that had recently been described for them in the relatively unfamiliar, passive voice (e.g., "the flower was watered by the girl"). Children frequently produced the passive voice, a form of syntactic priming. These responses were performed while children sometimes retained a visual-spatial or auditory-verbal working memory load to be recalled after sentence production but there was no effect of the load on syntactic priming. In a second experiment, after hearing passive-voice descriptions of pictures, instead of relying on priming, the children were directly asked to describe the same pictures by speaking in the same "special way." Surprisingly, under a load, children more often used the passive voice as they were instructed to do, but at the expense of producing additional semantic and grammatical errors (including some nonsensical renditions such as "the girl was watered by the flower"). We propose that working memory, when available, is used to impose a quality-control process whereby the semantic fidelity of the response to the stimulus picture is preserved, here at the expense of disregarding the experimental instruction to reproduce the passive voice.