This week I wrote about a new book, Conflicted Copy by Sam Riviere, which uses AI to generate poems. The review appeared in Prospect magazine’s newsletter, The Culture, which — like all the best newsletters — comes out once a week. You can sign up here: https://www.prospectmagazine.co.uk/culture/62107/sign-up-now-to-receive-our-newsletter-the-culture
Last year, amid a flurry of clickbait about how the newly-launched ChatGPT could write passably bad poetry, I expressed scepticism that AI would be able to “improve” in this area. The machine-learning process known as “smoothing”, which Natural Language Processing models use to refine their prediction of word combinations, seemed to be the opposite of what we look for in poetry. You can read the whole post here:
Notwithstanding the fact that he has now written a book of poetry using Natural Language Processing, I suspect Sam Riviere feels similarly about the poetic limits of AI. His no-prisoners novel about the literary world, Dead Souls (2021), opens with a publishing industry in crisis due to the discovery of “fixed books” — old works that have been recycled by changing a few names. Editorial heads begin to roll (“for weeks, it rained heads”), readers become disillusioned, and then the unthinkable happens: people start buying books of poetry,
as if they harboured some innate form of truth that works of fiction, or drama, or even biography and autobiography, were unable to convey. Faith in these popular genres had been eroded to such a point, I reminded myself, that even poetry had become preferable to them. In this assumption the reading public was of course gravely mistaken […] I state this as an enthusiast of the form.
Irony and absurdity aside, there is wisdom in this fable. The idea that poetry is not “innately true” is as old as Plato dismissing Homer as an imitator:
For as a painter by a knowledge of figure and colour can paint a cobbler without any practice in cobbling, so the poet can delineate any art in the colours of language […] Imitation, then, is devoid of knowledge, being only a kind of play or sport.
But one counter-argument to Plato is that what a poet knows is language — words understood as an aesthetic medium in a way that can’t be imitated by a computer, because some of the poet’s most important methods are intuitive and irrational.
Recently, I’ve been revisiting Don Paterson’s essay on poetic technique from 2007, “The Lyric Principle”, which advocates the “free anagrammatizing” of phonemes (e.g. “leaf mould” into “dead failed mode”) as a way of completing a line of verse:
Take your musical bearings from the words on either side of the gap, and see if that input alone will lead you to the word. This is, of course, putting absolute trust in the phonosemantic principle. What’s strange is how often it works — even when, and perhaps especially when, the sense is quite different from that which you intended.
The variables at play here have to be weighed between ear, mouth and eye, and then decided by an imagination willing to choose a counter-intuitive meaning — none of which ChatGPT actually has. Take this line from Thomas Hardy’s “The Voice”, which I quoted in my earlier post on AI as an example of the kind of poetry it couldn’t write:
Wind oozing thin through the thorn from norward
“Oozing”, with its connotations of thick liquid, is not obviously the “right” word here. But one of the things that makes it such a memorably expressed perception is the way that both sound and meaning flip immediately into “thin through”, which itself calls back to “Wind oo-”. And then of course there is “thorn from norward”, making the whole line is a wonder of rolling sounds.
This morning, I took Hardy’s line to ChatGPT, removed “ooz-”, and asked it to fill in the present participle describing the action of the wind. It suggested “whistling”, “rustling” and “murmuring” — but when explicitly prompted only to use the sounds already available on either side, it couldn’t do it, coming up with “sighing” and (when I queried this) “whistling” and “hissing”:
One thing that this highlights, of course, is Hardy’s inspired decision to avoid the mimetic “thin” sibilance of “s” entirely in the line, instead using the thicker, voiced “z”, which is closer in mouthfeel to the voiced “th” we meet the middle of the line (“through the”). Not only can ChatGPT not hear words in this way, but when challenged to do so it will simply start lying — that is, imitating the kind of thing it seems right to say, even if that means claiming there is an “s” sound where none exists:
As I suggest in my review below, Riviere’s poetic intervention in Conflicted Copy seems to have been a kind of “fine-tuning” of AI text. To get this raw material, though, he first mined the computer’s capacity for fluent untruthfulness, as it pretends to have a consciousness able to remember both a life and the difference between words and images:
The windows opened on a warm October day and
there on the road below glinted the first spark of the new
century. My mind could instantly process that image, and
with the memory being so vivid it took no more than a few
seconds to see how beautiful those words above had been.
(“Darken Dogs”)
The poems in the book, according to the author’s note, were made between December 2020 and January 2021. For half of this period of composition, everybody in the UK was living in lockdown, with many of us spending a lot of time disembodied online. The neural network generating the text was, of course, oblivious to this. But I wonder if — via their human improver — lockdown nevertheless influenced the way these poems foreground the unreality of physical life. Take, for example, “True Mode”, which begins:
Now you’re done with
the usual paper and ink
(unless this sounds like a
bad idea), you’re free to
wander out of the study
room into the open air
and ends:
a vast river flowing
below you in the same
way you just read about
in your notebook.
In my review below, I explain why I think that Conflicted Copy captures a specific moment in the evolution of AI as a literary phenomenon, and uses the technological limitations of that moment as a kind of formal constraint. But the more I think about it, the more I hear it as haunted by the limitations of its own human moment too.
Sam Riviere, Conflicted Copy (Faber and Faber, £12.99)
Is Sam Riviere’s fourth poetry collection, Conflicted Copy, to be trusted? Its first word is “True”. But its second word, oddly, is “PDF”. And together these form the title of its first poem:
A picture-book filled entirely
with text, a soft blue font, white
ink and yellow borders, a simple
grid, a few lines on each page — a
gentle voice, singing as it speaks.
Is this a self-description? It opens a paperback with a cream, yellow and blue cover, published by Faber & Faber, home of such famous poets as T.S. Eliot, Sylvia Plath and Seamus Heaney—a publisher, that is, with high standards. Final proofs sent to print, approved by an editor, would be a “true PDF”.
A conflicted copy, however, occurs when a document is edited by more than one user. And someone seems to have introduced disagreements here. How can a picture-book be “filled entirely / with text” and yet also only have “a few lines on each page”? How can a font be blue if the ink is white? This is not the sleeves-up realism of a Heaney, opening his first book with the “clean rasping sound / When the spade sinks into gravelly ground”. Someone owes us an explanation.
Keep reading with a 7-day free trial
Subscribe to Some Flowers Soon to keep reading this post and get 7 days of free access to the full post archives.