A generative model learns by being shown a large collection of real molecules and adjusting itself until those molecules look likely under its distribution. What it absorbs is not a memory of individual compounds but the local patterns that recur across the collection: which rings appear, which functional groups attach to which scaffolds, which linkers connect which fragments, how often a particular arrangement shows up. Those recurring pieces are the model's vocabulary.
Generation then works by composition. The model assembles a structure piece by piece, choosing each next fragment according to the probabilities it learned. Because the choices are made locally and independently, the model can place fragments together in a combination that no training molecule contained. This is the origin of novelty: a molecule is new when the combination is new, even though every individual piece of it was seen during training. Nothing in the model requires the whole assembly to have been observed before.
The same mechanism explains the failure mode. Local plausibility does not add up to global plausibility. A model can join fragments that are each perfectly ordinary into an arrangement that is strained, unstable, or simply not useful — the pieces are familiar, the whole is not. Novelty and quality are therefore separate properties, and a generated structure has to be checked rather than trusted.