We used cross-modal generative AI models, which rely on the Contrastive Language-Image Pretraining (CLIP) encoder, to generate portraits of fictional characters based on their names. We then studied to what extent image generation captures names' gender and age connotations when information from linguistic distribution is rich and informative (talking names, e.g., Bolt), present but possibly uninformative (real names, e.g., John), and absent (made-up names, e.g., Arobynn). Three pre-trained Computer Vision classifiers for each attribute ex- hibit reliable agreement in classifying generated images, also for made-up names. We further show a robust correlation between the classifiers' confidence in detecting an attribute and the ratings provided by participants in an online survey about how suitable each name is for characters bearing a cer- tain attribute. These models and their learning strategies can shed light on mechanisms that support human learning of non- arbitrary form-meaning mappings.