Ask a large language model for a sonnet about grief, and within a few seconds you'll have fourteen lines, competently metered, imagery in roughly the right places, an ending that gestures at consolation without overreaching. It will not embarrass you if you read it aloud at a memorial. It is also, I want to argue, not a poem in any sense that should matter to a reader or a publisher — not because it lacks technical competence, which it very often has, but because the competence is the only thing it has, and a poem needs more from its maker than competence.
What Language Models Are Actually Doing
It's worth being precise about the mechanism, because a lot of the anxiety and a lot of the hype both rest on a misunderstanding of it. A language model produces text by predicting, one likely next word at a time, what a plausible continuation of everything that came before it would look like, trained on an enormous corpus of human writing. When it produces a sonnet about grief, it is not drawing on a memory of loss. It is drawing on the statistical shape of ten thousand sonnets about grief that other people wrote out of actual loss, and interpolating a new one that resembles them. That's a genuinely impressive technical feat. It is a fundamentally different act from what a poet is doing when they sit down to write about a specific dead father, on a specific morning, with a specific unresolved argument still hanging between them.
The Thing a Poem Needs That a Model Doesn't Have
What a poem asks of its writer, at minimum, is that something be at stake for them in the writing of it — that the poem be an attempt to work out something the poet did not already fully understand before starting. That's why a bad poem written by a grieving person can still be worth reading in a way that a technically flawless generated one usually isn't: the bad poem is evidence of an actual mind working against an actual difficulty. A model has no difficulty to work against. It has no morning after the funeral. It cannot be wrong about how it feels, because it doesn't feel anything to be wrong about. What it produces is the average shape of how humans have written about feeling something, which turns out to be a genuinely different object from a specific human writing about a specific feeling, however similar the two might look on the page.
A model has no morning after the funeral. It cannot be wrong about how it feels, because it doesn't feel anything to be wrong about.
Where This Leaves Workshops, and Readers
This is already changing how workshops and editors have to read. A poem that once would have been judged purely on the page now sometimes has to be judged partly on whether a human actually wrote it, which is a strange and slightly sad new layer of scrutiny to add to an art that used to only ask whether the writing was good. I don't think that scrutiny is going away, and I don't think it should. If a reader's experience of a poem depends, even partly, on trusting that a person meant it — and I think for most readers it does — then a poem's origin isn't a footnote to its meaning. It's part of the meaning. The same fourteen lines mean something different depending on whether a person lived through the thing they describe or a model interpolated the shape of ten thousand people having lived through it.
Why We Don't Publish It
This is also, plainly, why we don't accept AI-assisted work from contributors to The Poetry Library, and why every contributor agreement here says so explicitly. It isn't a purity test, and it isn't a claim that the technology is without legitimate uses elsewhere. It's a much narrower claim: that the specific thing we're asking a contributor for — a mind actually working against a specific difficulty in a specific poem — is the one thing a model cannot supply, no matter how fluent the sentences it produces. You can automate the sound of thought. You cannot automate the thing that made the thought worth having. Until that changes, if it ever does, we'll keep asking for the harder, slower, human version, even when the fast one reads almost as well on the page. Almost is the whole argument.
