There exists a fundamental paradox in linguistic cognition. Experiments show consistent sound-symbolic biases in people's processing of artificial words, yet the biases are not manifest in the structure of real words. To address this paradox, we designed an experiment to test the magnitude and source of these biases. Participants were tasked with matching nonsense words to novel object forms. One group was implicitly taught a matching rule congruent with biases reported previously, while a second group was taught a rule incongruent with this bias. In test trials, participants in the congruent condition performed only modestly but significantly better than chance and better than participants in the incongruent condition who performed at chance. These outcomes indicate the processing bias is real but weak and reflects an inherent learning bias. We discuss implications for language learning and transmission, considering the functional value of non-arbitrariness in language structure and underlying neurocognitive mechanisms.