I've noticed as a kid, that strawberry flavoured candy doesn't actually taste like strawberries. They are clearly and collectively recognisable as strawberry candies, but that's just pattern matching and conditioning on wording. The flavour has not much to do with actual strawberries, even the sweetness is vastly exaggerated. The synthetic aroma is much less complex, as the author noted. We just fell into the habit (or trap) of using the same word for both flavours.

On the other hand i'm wondering if that's just an implementation detail. A temporary imperfection in simulating the real thing due to constraints in (chemical) engineering and cost, not a hard limit.

Neural Networks are universal function approximators. Throw enough resources at them and they will mimic the most complex function to an arbitrary level of detail.