Form aware AI poetry generation in 2026, is the bland mush critique still valid - any thoughts

[Highland Builder]

There is a piece making the rounds arguing that the AI poetry is bland mush critique is fair for prompt and pray output but not for current form aware generation. The claim is that 2026 systems enforce structural rules at generation time so sonnets actually land at fourteen lines with a proper turn, villanelles maintain their refrains, ghazals hold their radif.

I tried this last week with a few of the newer models and the structural part is genuinely true. Free verse still reads like a competent MFA workshop submission with no specific human inside it, but the formal verse output is at minimum technically correct in a way it absolutely was not eighteen months ago.

Which raises an uncomfortable question. If a sonnet is technically perfect, metrically clean, holds its volta, but has no actual person behind the speaker, is it a poem or just a poem shaped object. I genuinely do not know where I land on this anymore. The Nature paper from 2024 showed non experts cannot tell the difference, the expert critique has now narrowed to a vibe that nobody can quite articulate

[TheRock]

A poem shaped object is exactly the right phrase, that is what AI formal verse always is

[Beth3.0]

But what makes a sonnet a sonnet other than the form, isnt the form the whole game

[Forge37]

No, the form is the constraint, the poem is what you do under the constraint, those are different things

[DeepInlet]

I tried generating villanelles last week and you are right they are technically clean but the refrain never earns the second appearance the way a good villanelle does

[StringTheory51]

The earning part is the entire point and the model has no concept of earning anything

[FrostCandle]

Honestly I think the formal verse community is in more danger from AI than the free verse community

[Chris_50]

Backwards, free verse is in more danger because the lack of constraints makes the genericness less detectable

[Craig]

What about hybrid forms, anyone tried these with current models

[One-One-Five]

Tried with Anne Carson style hybrid stuff and it completely falls apart, you can see the system has no internalised reason for the formal choices

[GoldbergFan_X]

Which actually makes me feel better about my own writing weirdly

[TommyB_20]

The fact that we are having this conversation in 2026 is itself the answer to whether this matters