[Jerrrrrrry]

The model is likely to had already seen the exact phrase from its last iteration. Adding variation generalizes the inference away from over-trained quotes.

Every model has the model before it, and it's academic papers, in it's training data.

Changing the qualifiers pulls the inference far away from quoting over-trained data, and back to generalization.

I am sure it has picked up on this mesa-optimization along the way, especially if I can summarize it.

Wonder why it hasn't been more generally intelligent, yet.