The model is trained on lines, and the word 'pelican' and not much more.

The model does not 'understand' comprehensively the relationship between anatomy, dimensionality, etc..