This item is taken from PN Review 287, Volume 52 Number 3, January - February 2026.
Editorial
A friend challenged me: ‘Tell ChatGPT to write a poem in your style.’ I am so posthumous as a poet that I doubted ChatGPT had ingested and digested my gristles. But I asked it and, sure enough, in seconds it regurgitated – or gurgitated – a poem clearly using devices I favour, with my repetitions, my two caesura lines, my strained diction. The programme knew me. Or it knew about writing that it attached to a person with my name. What ChatGPT delivered was not quite parody, not quite poem. It was, however, recognisable; a casual acquaintance encountering it might have said, ‘He’s writing again, though repeating himself…’. I then told ChatGPT to write a 2,000-word introductory essay on the Roman poet Ennius, with footnotes. This it did, again plausibly, except the footnotes were all over the place. Students who use ChatGPT for essays are often snared on wayward footnotes… The programme is trying to repair this Achilles heel.
ChatGPT, at first, alarmed me. Working through the poetry and prose submissions that PN Review receives through its narrow ‘submission windows’ – over 900 in the rarely scheduled week – I wondered if any, or rather how many, were machine generated or assisted. The absence of handwriting, of Tipp-Ex, of crossings out, meant the only evidence I had to go on was the digital attachment without – despite the term digital – fingerprints.
But isn’t it the case that all of us who love and read poetry, and know much of it by heart, are ourselves generating our own poems in a way analogous to the programmes with which I was experimenting? The main difference was that a subjectivity, or a deliberate creative purpose, governs our output, not an external command fed into a machine. 1 The ‘anxiety of influence’ programmes are within us, distinctive to each of us as living, breathing readers; it is not something we access on our laptops or iphones. It is not play. Even when we joke or ironise, we are in earnest.
I am currently writing about poetry in Latin from Roman times to the present day (hence the Ennius example) and I’ve reflected on how Virgil, Horace and Ovid (among others) are Large Language Models (LLMs) in themselves. If you were a Renaissance writer of Latin poems (the Medieval poets had more fun, flirting with the demotic while drawing on the classics), you strove for a correctness authorised or underpinned by the great Latin writers. They themselves were programmed by Latin and Greek antecedents.
The later Latin poets were not using computers but pens and ink, and their work was being read by people who could appraise their success in its own terms. Before the work went to the printer there were actual thumb smudges and emendations, and then in proof (there were printers’ proofs) further changes could be introduced.
Admittedly, there are writing programmes where pupils are urged not to read too much for fear of marring their inborn originality. But that ‘originality’ is itself shaped by the clichés of the language surrounding them: no one who writes, or performs, a poem possesses an innocent language. Even William Blake and his supposedly illiterate wife Catherine Boucher, sitting naked in conversation in their garden, were steeped in the classics, gospels and Psalms.
The poetic ecosystem hasn’t radically changed – though the new tools introduced have a sort of mind of their own. They have nothing particular to say (like so many living, breathing poets), but they mimic thought and speech and they can chance upon compelling effects. I sift through the submissions. Is this tiny grain of gold real or fool’s gold? Is there a qualitative difference between the two? Am I judging the poem before me by its effect, by what it does? If the poem does something unusual, does it matter what – not who – composed it?
Poetry is a collaboration. Even the most original-seeming writers are influenced by formal language and carry within them their own LLMs. These language models empower and restrict what they can do. In a way AI tools are freer than the human being because their LLM is much larger and more diverse than any mere individual’s, however well-read that individual may be, and are therefore in a sense more at liberty. Not least because they needn’t say anything (unless it receives instructions to say something).
But that freedom is remote from the freedom poets enjoy. We have no evidence that enjoyment is part of the ChatGPT process, yet it is a transmissible ingredient of the achieved poem as of the poet’s experience of writing. So long as enjoyment remains at the heart of poetry and the other humanities, they are (for the time being) proof against technology and the ideologies that support and overstate its case.
When I delivered the text of my book The Novel: A Biography to my publisher, they fed it, without informing me, through a plagiarism detector, a precursor to the programmes that check student essays for borrowings and computer collaborations. My editor reported that plagiarism had occurred at one point in my book. I had quoted, without attribution, a review I myself had written. I had to alter the text. This was my first taste of what would be in store. In the twelve years since, the world has come to look different. How different is it, really? I’ll tell you when I have read through the December submissions. I may be some time.
ChatGPT, at first, alarmed me. Working through the poetry and prose submissions that PN Review receives through its narrow ‘submission windows’ – over 900 in the rarely scheduled week – I wondered if any, or rather how many, were machine generated or assisted. The absence of handwriting, of Tipp-Ex, of crossings out, meant the only evidence I had to go on was the digital attachment without – despite the term digital – fingerprints.
But isn’t it the case that all of us who love and read poetry, and know much of it by heart, are ourselves generating our own poems in a way analogous to the programmes with which I was experimenting? The main difference was that a subjectivity, or a deliberate creative purpose, governs our output, not an external command fed into a machine. 1 The ‘anxiety of influence’ programmes are within us, distinctive to each of us as living, breathing readers; it is not something we access on our laptops or iphones. It is not play. Even when we joke or ironise, we are in earnest.
I am currently writing about poetry in Latin from Roman times to the present day (hence the Ennius example) and I’ve reflected on how Virgil, Horace and Ovid (among others) are Large Language Models (LLMs) in themselves. If you were a Renaissance writer of Latin poems (the Medieval poets had more fun, flirting with the demotic while drawing on the classics), you strove for a correctness authorised or underpinned by the great Latin writers. They themselves were programmed by Latin and Greek antecedents.
The later Latin poets were not using computers but pens and ink, and their work was being read by people who could appraise their success in its own terms. Before the work went to the printer there were actual thumb smudges and emendations, and then in proof (there were printers’ proofs) further changes could be introduced.
Admittedly, there are writing programmes where pupils are urged not to read too much for fear of marring their inborn originality. But that ‘originality’ is itself shaped by the clichés of the language surrounding them: no one who writes, or performs, a poem possesses an innocent language. Even William Blake and his supposedly illiterate wife Catherine Boucher, sitting naked in conversation in their garden, were steeped in the classics, gospels and Psalms.
The poetic ecosystem hasn’t radically changed – though the new tools introduced have a sort of mind of their own. They have nothing particular to say (like so many living, breathing poets), but they mimic thought and speech and they can chance upon compelling effects. I sift through the submissions. Is this tiny grain of gold real or fool’s gold? Is there a qualitative difference between the two? Am I judging the poem before me by its effect, by what it does? If the poem does something unusual, does it matter what – not who – composed it?
Poetry is a collaboration. Even the most original-seeming writers are influenced by formal language and carry within them their own LLMs. These language models empower and restrict what they can do. In a way AI tools are freer than the human being because their LLM is much larger and more diverse than any mere individual’s, however well-read that individual may be, and are therefore in a sense more at liberty. Not least because they needn’t say anything (unless it receives instructions to say something).
But that freedom is remote from the freedom poets enjoy. We have no evidence that enjoyment is part of the ChatGPT process, yet it is a transmissible ingredient of the achieved poem as of the poet’s experience of writing. So long as enjoyment remains at the heart of poetry and the other humanities, they are (for the time being) proof against technology and the ideologies that support and overstate its case.
When I delivered the text of my book The Novel: A Biography to my publisher, they fed it, without informing me, through a plagiarism detector, a precursor to the programmes that check student essays for borrowings and computer collaborations. My editor reported that plagiarism had occurred at one point in my book. I had quoted, without attribution, a review I myself had written. I had to alter the text. This was my first taste of what would be in store. In the twelve years since, the world has come to look different. How different is it, really? I’ll tell you when I have read through the December submissions. I may be some time.
- (Though could a subjectivity be simulated? If OpenAI is now valued at US$500 billion as an industry, it would cost small change to fake it…)
This item is taken from PN Review 287, Volume 52 Number 3, January - February 2026.