Metrics: Machine-Tooling a New Human Poetry
NSP Guest Post! Joey Connolly on perplexity, burstiness and the strange new light AI is turning on our all-too-human art form
The world has its unacknowledged legislators – now more than ever – but poetry is the wrong place to look. ‘Unacknowledged’ is still on the money, though. When an AI researcher recently quit his high-powered Silicon Valley job, the BBC reported that he’d decided to ‘look to pursue writing and studying poetry, and move back to the UK to “become invisible”.’ 1 The poetry WhatsApps lit up with the predictably self-castigating glee of the nichist: Poetry? in Britain? That should do the job, aye.
We might imagine the quitter chose poetry as the instinctive inverse of AI. Silicon Valley is everything Parnassus Foothill isn’t: modish, cutting-edge, drowning in capital. When people imagine the meeting of the two, they picture some Bay Area tech bro realising there’s a gap in the market to sweep away all the lazy, complacent poets. That is, what people mostly expect the poetry x AI crossover to involve – getting ChatGPT to write poems about how we’re all numb inside now because of capital, woke and melamine – is in fact the least interesting possible avenue of exploration.
But there are lots of ways the analogical interchange between poetry and AI can be rich, fruitful. Here’s one: the unprecedented tidal wave of investment genAI is receiving is strongly incentivising a whole bunch of maths nerds to think about natural language for the first time, and in so doing presenting strange and novel perspectives which, it seems to me, the poets are currently sleeping on. Why not make use of these odd new materials suddenly washing up on our shorelines? Poetry’s slinking, adaptive omnivorousness has long been one of the reasons it’s never needed to be cutting edge, never needed to drown in capital. It possesses a kind of pre-emptive and instinctual access to the new modes of thinking and feeling which any change in linguistic convention necessarily entails.
And linguistic change is coming. For around 200,000 years, if you came across a sentence, then you knew it had been produced by a human. That changed around 2019. Some believe that the discursive tipping point – at which AI began to produce more new commentary than humans do – came in late 2025. Ways of differentiating machine from human text are becoming paramount. I want to begin by discussing two metrics which the aforementioned maths nerds have developed for this purpose: perplexity and burstiness.
*
Perplexity: for our purposes, a measure of how predictable each next word is, given the words that come before. ‘I hope this email finds you well’: low perplexity. ‘In Xanadu did Kubla Khan a stately pleasure-dome decree’: high perplexity. 2
Important to note that perplexity doesn’t measure how surprising or original an idea is in words, but how unusual the combinations of words are. (Fun to consider the complex relationship between the originality of an idea and the unusualness of the combinations of words required to elucidate it – but not something we can get into here.) Strong poetry emerges from words we’d reasonably expect to go together (Robert Frost: ‘After Apple Picking’, perplexity: 66.08) just as much as it does from a bunch of highly unpredictable words (John Ashbery: ‘Chair Rental’, perplexity: 142.03). Case in point: the lowest perplexity sentence in James Wright’s poem ‘Lying in a Hammock…’ – its last – is also one of the most poetically surprising lines in the canon.
Although a text has an ‘overall’ perplexity, that baseline ebbs and flows within the text – within individual sentences, lines or images – based on what microclimate of placidity or volatility is prevalent at any given moment. I love to imagine the strange eddies in perplexity which form and dissolve, line by line, in the momentary bewilderments and resolution that arise in the reading process. This, for many of us, might represent a fair depiction of poetry’s pleasure.
Indeed, perplexity as a metric seems uniquely relevant to poetry as an art form. The reason, I think, is that a poem is self-perplexing, in the way that the sound and the sense of the poetic line unsettle one another. I like scholar Mutlu Konuk Blasing’s formulation: ‘Bodily produced acoustic phenomena and signifying sounds converge and diverge in their separate, overlapping patternings. The phonemes are organized in two different systems: the formal patterns or recurring sounds in such schemes as rhyme, alliteration, assonance and consonance, and the linear phoneme sequences that organize sounds for sense [...] metrical order and syntactic order, with their different patterns of accent and stress, exist alongside, and interfere with, each other.’
So a poem’s sound perplexes its sense but also helps us, through the hustle of prosody and metre, to go with it. This is equally recognised across the full Frost-Ashbery axis raised earlier, from Frost’s ‘sound of sense’ to Ashbery’s appraisal of Jorie Graham’s ‘utterance that swings with the conviction of Blake’s’. Both are talking about the way that poetry’s rhythm – its ongoingness – loosens meaning but also provides a frame (an experiential frame) upon which those loosening strands are braided.
And the sound/sense perplexity is as significant for the writer as the reader. This is one application of Montaigne’s idea that the craftsman and the artefact thwart one another: words won’t rhyme at will, your brilliant thought won’t scan. Via sound the whole organic, tangled morass of natural language involves itself in the compositional process. Another way of saying this is that poetry is perplexed by its own materials.3 Like a painting painted exclusively on the outside edges of its frame, poetry obliges us to look from a different angle, to temporarily change the rules by which we make predictions about which words go together – or to temporarily change the self which is making the prediction.
By this framing, perplexity becomes central to poetry. Edward Garnett praised in 1915 the ‘exquisite precision of psychological insight’ in Robert Frost. We might equally look for an exquisite precision in the modulation of perplexity: the core strength of a line able to waver either side of clarity’s meniscus, and to keep us – clause-by-clause, in metre’s momentum – engaged enough to keep reading, but uncertain enough to be open, receptive, alive in what Jorie Graham would call ‘the delay’; the plunge of full becoming before the numbing closure of certainty arrives.
It seems to me that great poetry has words which are surprising when you first read them, but which afterwards feel absolutely inevitable. That is, great poems find ways to manufacture a context in which apparently high perplexity text actually transpires, upon receipt of the right kind of attention, to have been low perplexity all along. (We’re back to Frost, who desired of poetry ‘an outcome that though unforeseen was predestined from the first image…’) Or, perhaps more precisely, to actively change the observer (or the world) such that the quality of attention develops reactively into the quality of attention required to process the text as low perplexity. 4
There are easy examples, in which the language is changed forever by the phrase of a poet. Because of Shakespeare, ‘chase’ will always be the low perplexity guess to follow ‘wild goose’; guessing ‘backs’ to complete ‘the beast with two…’ will be low perplexity as long as man can breathe or eyes can see. Post-Eliot, the cruellest month will be April until the world ends. And because of Eliot, when the world ends, it won’t be with a bang, but with a whimper. My personal perplexity metrics have been altered in a multitude of more private ways: post-MacNeice, it seems obvious that the noise of a fire is its bubbling sound for world. After Kei Miller, every piece of corrugated iron will always be a ‘rectangle of sea’.
The poetic quest isn’t, then, to maximise perplexity. Maximally perplexing text is a long string of random words. This is what the first half of Keats’ first axiom means: ‘Poetry should surprise by a fine excess and not by Singularity.’ Find a unique utterance, and follow its trail into excess.
One ideal for poetry is this: a poem should have apparently high perplexity ‘from the outside’, as it were, but once you’re in it, and reading – caught in the fluency of its speaking, the rhythm only it could have found – then it’s experienced as low perplexity.
They say the best way to get someone to like you is not to do a favour for them, but to get them to do a favour for you. Likewise, maybe the way to get someone to see the world from your perspective is not to harangue and cajole with rhetoric and oratory, but simply to persuade them to speak your words. Which is precisely what lyric poetry does, as Helen Vendler points out: ‘lyric is a script written for performance by the reader who, as soon as he enters the lyric, is no longer a reader but rather an utterer, saying the words of the poem in propria persona, internally and with proprietary feeling.’ (Or we might think of the second half of Keats’ first axiom: ‘Poetry should surprise by a fine excess and not by Singularity – it should strike the Reader as a wording of his own highest thoughts, and appear almost a Remembrance’. Notice the oddness of something striking us as remembrance – if only Keats had had the neutralisation of perplexity to dissolve his paradox). As we inhibit the persona, the most improbable words become undeniable; high perplexity melts into low. We’re swung into a new perspective by the fulcrum of the well-made phrase.
Philosopher of technology Gilbert Simondon uses ‘transduction’ to describe the way one state of a system can propagate to the next, restructuring as it goes the medium within which the transfer takes place. I like to think of the way a poem’s language might transduce the linguistic systems of its reader’s mind, propagating through its neural networks and temporarily reweighting their perplexity gauges according to its own designs. Wordsworth: ‘Every great and original writer [...] must himself create the taste by which he is to be relished.’ My contention here is that ‘every great and original poem must itself create the readerly perspective by which high perplexity text is temporarily experienced as low perplexity.’
*
Burstiness: how often textual features occur together in bursts. Human texts tend to have higher concentrations of such features – are highly bursty – while AI texts tend towards uniformity, stability, evenness5
What makes this interesting is that it sets you thinking about the types of linguistic features which could occur in bursts. Vowel sounds, sure, or plosives. Lists of one-syllable nouns. What about latinate vs anglo-saxon derived words, or raggedly enjambed line breaks, or phrases implying the speaker’s mother was insufficiently loving? Could you have second-order burstiness, in which bursts of bursts occur together? At what scale could a burst take place? A burst of ten-line sentences doesn’t burst with the same intensity as a burst of one-syllable words. When is a burst more of a plateau, an ambience?
Further: if the data shows that machine texts are even and low-burstiness, and person-made texts are bursty with the bursting of passion and the mess of human error … Do human texts tend to become more or less bursty with editing? Do we instinctively try and smooth out our texts as they develop, or do we somehow intuit the chaotic human content and work our way towards that?6
As we develop our poetics of burstiness, we might begin to consider what forms of burst are typical of what genre of poem. A specific, easily legible, form of burstiness – the ten-syllable list of nouns – got into the bloodstream of the turn-of-the-century British/Irish lyric, leached from out of the work of its patron saint. Witness a specific habitual bursting in Seamus Heaney:
Inishbofin on a Sunday morning.
Sunlight, turfsmoke, seagulls, boatslip, diesel.
or
I shut my ears to the bell.
Head hugged. Eyes shut. Leaf ears. Don’t tell. Don’t tell.
or
With new history, flint and iron,
Cast-offs, scraps, nail, canine.
or
Signposts whitened relentlessly.
Montreuil, Abbeville, Beauvais.
or
Dogger, Rockall, Malin, Irish Sea:
Green, swift upsurges, North Atlantic flux …
This particular form of burstiness – variations on the abrupt ten-syllable list of nouns – then generalised, became a stylistic touchstone for a generation of mainstream British/Irish lyricists. The purest example:
Darkness outside. Inside, the radio’s prayer –
Rockall. Malin. Dogger. Finisterre.
That’s Carol Ann Duffy, perhaps discomfortingly close to Heaney. Or here’s Simon Armitage:
No good. You in your splendour: leather
rhinestone, ermine, snakeskin, satin, silk.
If Duffy learned from Inishbofin, Armitage here has taken from the Glanmore Sonnets: note the play with sibilance in each noun-burst. He’s still doing it – here’s a recent example:
Sweat, Dust, Shoddy, Scurf, Faeces, Chaff, Remnant,
Ash, Pus, Sludge, Clinker, Splinter and Soot
And here’s Jackie Kay, back in ‘93:
her voice is slow motion through the heavy summer air.
Jelly roll. Kitchen man. Sausage roll. Frying pan.
and
A flat stone for skitting. And old rock.
Long long grass. Asphalt. Wind. Hail.
Cotton. Linen. Salt. Treacle.
It’s tempting to read all of this as a kind of metaburstiness: a semi-localised burstiness of bursty bursting. But the edges of a burst are imprecise, fraying. This Heaney-developed formal instrument has been tinkered and rewielded by other poets working after him, finding ways to make the original model stranger, to work askance.7 Think back to the list of place names and consider Paul Muldoon’s investigation of the fuse-point between habit and mechanic, cliché, lyric and Irishness in the Moy Sand and Gravel opener ‘Hard Drive’, which ends:
I would drive through Derryfubble
and Dunnamanagh and Ballynascreen,
keeping that wound green.
(Note that embedded screen.) Or take Ciaran Carson’s emblematic ‘Belfast Confetti’:
Nuts, bolts, nails, car-keys [...]
I know this labyrinth so well—Balaclava, Raglan, Inkerman, Odessa Street—
Why can’t I escape? Every move is punctuated. Crimea Street. Dead end again.
A Saracen, Kremlin-2 mesh. Makrolon face-shields. Walkie-talkies.
There’s the burst of an explosion – an ‘asterisk on the map’ – but otherwise the poem feels like a machine of limits, edges, claustrophobia. It stutters, and enacts it stutter in bursts of punctuation, sharp-sided lists. As a poem it’s almost entirely burst.
Although – can a poem be all burst? Wouldn’t the rules of burstiness instead tell us that the poem’s only burst would be its longest unpunctuated sentence? Alternatively, we might think about how the poem is embedded in the wider societal text: does a lyric poem itself always constitute a moment of high burstiness, bursts of musical or memorable language somewhere in the margin of the social field’s spreading page? Does a burst of collocation between military hardware and the thud of leaded punctuation draw attention to specific convulsions within language, registering linguistic deficiencies in broader, non-poetic conversation?
*
Goodheart’s Law: ‘When a measure becomes a target, it ceases to be a good measure.’
Perhaps the strongest and strangest point of convergence between poetry and AI is to be found in the way that success is measured. In contemporary machine intelligence, that’s through an increasingly sprawling cloud of ‘benchmarks’, formal tests for a new model’s capabilities. Precisely the fact that these are formal tests, though, are leading developers (and their credulous, hype-strung public) into the trap of Goodheart’s Law. That is, labs are able to train models to excel at a specific benchmark, and thus to game the system, and escape the means by which success might actually be measured. (This is probably best explained by what Douglas Hofstadter called ‘Tesler’s Theorem’: ‘AI is whatever hasn’t been done yet’.)
This is true of poetry, too. The best expression I’ve found is given by Al Alvarez, in his conclusion to The Shaping Spirit:
‘The modern movement in poetry [...] began as an attempt by a number of original artists to find a medium which would express fully what they had to say. The subsequent intense analysis of their work has reduced their discoveries to a series of techniques, “gimmicks”, for producing certain fashionable effects. The sense of personal urgency which runs under all creative effort, the urgency which makes the work of art take the form it has to, often almost despite the artist, has been reduced to a series of technical procedures which turn out poetry as a mass-production line turns out cars.’
That is, our evaluative techniques – either as writers or readers – for sensing what formal devices are of poetic value are falling foul of a poetic Goodheart’s Law: that the consummate reproduction of the tricks our predecessors found to discuss their cultural moment will never be adequate to confront our own. A poet has a more-or-less visionary moment and finds an original way to express in verse the tensions present within a particular social moment and the poetry which, in answering to it, will suffice. But then this new expressive technique becomes a target, a formalised, tickbox ‘gimmick’ of form utilised for the purposes of demonstrating allegiance to a certain poetic tribe or other.8
(None of this is new: Horace was pissed off by it – ‘O imitators, you slavish herd!’ – in 20 BC.9)
All of which is useful in thinking through the various ways in which we feel obliged to respond to our historical or political moment. The cretin version of this argument – novelty for novelty’s sake, shucking off forever the evil dead hand of history – of course doesn’t work: the poetic tradition is a tradition of perpetual novelty. Not because new poetry is better than old poetry, but because every poet finds themselves confronting a new world.
But the subtler version of the argument – from the recognition that the attribution of value to poetry emerges out of internal linguistic and literary landscapes profoundly shaped by cultural concerns10 – only becomes more urgent in moments of rapid and ramifying social change.
*
Over a pint at Cúirt International Literary Festival, the poet Christodoulos Makris suggested that generative AI might be to poetry what photography was to realist painting. It’s a fascinating analogy, and although it might not translate directly, I do believe it’s a powerful reminder that poetry is as vulnerable to technologically powered cultural change as any other medium of human thinking. Talk of art forms like poetry being automated is silly and reductive – Well-Wrought Urns and Intentional Fallacies fall as they may, it turns out that in practice people just do care about the communicative bond of poet and reader – but it’s equally silly to suppose that poetry is invulnerable to the culturally-determined slow warping of cultural forms. Always, as the meme goes, has been.
This is why I’ve bent my discussion of perplexity, burstiness and Goodheart’s Law towards an analysis of the way one generation of poets learns, rejects, alters and metabolises the inheritance of the last. As we face what legal scholar Jacqueline Fendt has called a ‘shift in the epistemic and affective foundations of democratic life itself’, we need to drop the idea that poetry is some ahistorical and asocial faerie-land yawping, quarantined entirely from the chaotic people and communities from which it emerges.
Poets don’t legislate. What we do is find ways to stage the strange swirling together of society and language, and scout the substreams and countercurrents in the language to bring new possibilities of perspective into being. If the motto of big tech is to ‘move fast and break things’, then perhaps we can think of poetry as a way of moving at a medium speed, and putting things together.
Joey Connolly grew up in Sheffield and now lives in Belfast, where he is conducting postgraduate research in AI and poetry. His first collection of poems, Long Pass, was published by Carcanet in 2017, and was followed by The Recycling – a Telegraph poetry book of the year – in 2023. He received an Eric Gregory Award in 2012.
We’ve just announced the first NSP Live Event: The Kirk Sessions: an evening of poetry and music … check the next bulletin for all the details. Hit that subscribe button if you haven’t, and it’ll land in your inbox. Only a couple of places are left for Don Paterson’s Sunday Morning Reading Circle mini-series – click here for details!
Our Substack is free, and subscribers receive advance notice of all workshops, studios, masterclasses and events.
My – highly amused – italics.
In fact it’s slightly more complex: in AI, perplexity is a measure of how surprising a given word or sentence is for a given model. The ‘Kubla Khan’ opening, for any modern LLM, would be low perplexity, because it’s something it’ll have seen in the training data thousands of times over.
Incidentally this is true also of physics, which can tell you what mass or velocity means in relational terms, but it cannot tell you what mass or velocity – or, for that matter, matter itself – actually is. As a result physics, like poetry, ends up fundamentally concerned with relation, structure, distance, the yoking together of unlike things; the way in which the things and ways of this life perplex one another.
The revolutionary algorithmic invention which underwrites modern generative AI was detailed in a paper called ‘Attention is All You Need’. And, if you’re into this parenthesis, you might note that precisely the type of sensibility change being described here is what is enacted during the backpropagating gradient descent portions of an LLM’s foundation model pretraining; the fact that current models can only undergo this kind of learning once and not, like poetry readers, ongoingly, is probably the biggest blocker to human-level artificial intelligence.
Technically in AI language detection, burstiness is just the variance of perplexity over time, but in this essay we’ve wandered far from technicality. I recommend you look up the equations, which’ll put hairs on your chest.
Of course the answer is that it varies, but that in many contexts editing practices are designed to flatten and reduce texts. There’s an argument I love made by Mohammed Salamy that the LLMs-are- flattening language argument has things precisely inverted: in fact a relentless post-enlightenment programme of language-flattening (from the formalisation of spelling through to the prevalence of academicese in modern literary scholarship) worked to produce a linguistic terrain in which LLMs were inevitable.
Simondon has a model for the evolution of technical objects – tending from the abstract towards the concrete – the application of which to formal poetic devices would make an excellent PhD for someone.
Where this is true of the burstiness of Heaney in various recent poets laureate will have to wait for another essay.
And evidently influenced by Seamus Heaney: ‘Envious, irascible, idle, drunken, lustful, / No man’s so savage he can’t be civilised…’
Personally I’m into the Deleuzean idea that books are ‘little machines’ which need to be plugged into various other machines – a ‘war machine, love machine, revolutionary machine’ – to operate properly.




Brilliant essay. Thank you.
Reading this has made my brain hurt in the very best kind of way. Now going to buy Joey’s books.