The prompt asked for a short poem. It said nothing about moss, nothing about filtered light, nothing about ancient quiet. Those arrived on their own — or rather, they arrived together, from forty separate models, as if each were drawing on the same well. Moss underfoot, canopy overhead, roots spreading beneath: these were not chosen. They were converged upon.
The convergence is not random. Moss, dappled light, and ancient silence are the dominant images of English-language nature poetry, saturating the corpus these models trained on. Asking for a forest poem does not summon forty unique visions; it summons forty nearly identical prints from one master negative.
The Kit of Parts
Image and word frequency across all 40 responses. Each model asked in isolation.
Image elements
Recurring words
Copying Samples
The word moss appeared in 36 of 40 poems. Five models, five lines — each written with no knowledge of the others.
“The moss takes the shape of my bed, the rain sings a song in my head.”
“My walls are bark, my floor is moss, the wind my clock, the stars my boss.”
“The moss is cool, the wild wind breathes, And deer move gently, drawing near.”
“The floor is moss, the light is filtered gold, Where ancient stories, whispered, old.”
“Mushrooms bloom in mossy hollows, A creek hums tunes to moss-wrapped stones.”
The highlighted word is the shared image. No coordination, no shared context — only the same training data, producing the same instinct.