This is a good example of something that I feel like I need to drill at a bit more. I'm pretty sure that this isn't an unexpected behavior or an overfitting of the training data. Rather, given the niche question of "what time zone does this tiny community use?" one relatively successful article in a satirical paper should have an outsized impact on the statistical patterns surrounding those words, and since as far as the model is concerned there is no referent to check against this kind of thing should be expected to keep coming up when specific topics or phrases come up near each other in relatively novel ways. The smaller number of examples gives each one a larger impact on the overall pattern, so it should be entirely unsurprising that one satirical example "poisons" the output this cleanly.
Assuming this is the case, I wonder if it's possible to weaponize it by identifying tokens with low overall reference counts that could be expanded with minimal investment of time. Sort of like Google bombing.
If it's regurgitating memorized snippets of human-created training data it's arguably not even artificial.