In their new collection Machinations, Kinneson Lalor and JP Seabright take inspiration from Alan Turing and the world of artificial intelligence, creating poems that are conversations not only between two poets, but between poet and machine. Here, they share their experiences when working on the collection, along with some of the resulting poems.
Machinations is a poetry collaboration that started with a conversation, not about shared interests, but about spaces we were hesitant to go as writers. One of us has a background in mathematics and computational physics, but thought that being a writer meant turning away from these disciplines. The other finds it difficult to process numbers, and so was curious to explore concepts they didn’t feel they had natural access to. Deciding to focus on the work of Alan Turing came as a natural consequence of these conversations, not only because his work touched on many of the themes we were interested in, from binary arithmetic to exploring machine-assisted creativity to surveillance, but because his personal life – his queerness, and consequent conviction and death – spoke to us as well.
The collection is not just a collaboration between two writers. Inspired by Turing’s work with artificial intelligence, it’s also a collaboration between humans and machines. We have used various algorithms to create poems, such as cut-up and rub-out generators, OpenAI’s GPT3 (a language processing model that uses machine learning to generate human-like text from a prompt), and image processing software created just for this collection. For example, the poem ‘strawberries’ was made using a cut-up and rub-out generator applied to statements in Turing’s paper Computing Machinery and Intelligence (Mind, vol. LIX, Issue 236, October 1950, p433-460) about ‘various disabilities’ a computer might have in comparison to a human.
At the centre of this collection is a question Turing himself posed as a possible limitation of a computer: can a machine write a poem? In exploring this question, we wanted to also incorporate the wider breadth of Turing’s work, and throughout, we’ve used his personal letters, drafts, sketches, rough working, and papers to create the poems. Some of the best examples of this in the collection are the poems that use especially-created image-to-poem algorithms. The code creates poems from black and white images via an evolving binary-phoneme dictionary where unique phonemes are assigned a seven-bit number. For example, 0101000—which visually represents two occurrences of a black pixel followed by a white pixel then three black pixels—was assigned to /səns/, as in sense, nonsense, absence etc. To incorporate the spread of Turing’s intellectual endeavours, various models from Turing’s own papers were implemented, and a visual snapshot of the output was processed to produce a poem, or vice versa. For example, the image, and accompanying poem, in ‘It is found experimentally that the axis is in some definite direction’ was produced by implementing the reaction-diffusion equations in Turing’s paper The Chemical Basis of Morphogenesis (Philosophical Transactions of the Royal Society of London. Series B, Biological Sciences, Vol. 237, No. 641. (Aug. 14, 1952), pp. 37-72). In the paper, Turing proposed how the patterns we see naturally occurring in nature, such as spots on a leopard or the markings on a giant pufferfish, might arise from more uniform states via a series of interactions of two diffusible substances. It’s easy to find the metaphors and see why a man prosecuted for the differences he was born with might be drawn to such work. In using software to reproduce his work and then produce a poem from the output, we were looking for the man behind the maths by using his own tools.
In the algorithm, to begin with, the dictionary is empty, but with each poem it grows. Eventually the poet is left with fewer gaps to fill as the algorithm starts stringing together phonemes that aren’t actual words, and the poet has less and less control over the output. Finally all that’s left to the poet as a poetic tool is whitespace. A snapshot in time is taken, processed, and a string of phonemes are output. The poet’s job is then to find the best placement for these phonemes, whether that’s how they appear on the page and interact with the image, or whether neighbouring phonemes can be placed together to create any sort of poetic sense.
Each poem is linked to Turing’s work and unique as a result. The moment the snapshot is taken is random and likely unreproducible, and because some of the simulations are different every time, such as the random walk simulation, you are taking a truly unique image to produce a unique poem. If we did the process again, not only would we get a different poem, it would also add different phonemes to the binary-phoneme dictionary. The algorithm changes as the input changes, and hence so does the poem and all subsequent poems that use that dictionary. It’s an important lesson in utilising machines for the dissemination of ideas. Just ask OpenAI’s current machine-learning tool to write a sonnet about your wife. Drawing on its access to the entire universe of the internet, it produces something about how great she is at cooking and being a mother, regardless of whether you have children or what her actual skills in the kitchen are. It will, however, get the sonnet form perfect.
This isn’t the only algorithm used in the collection that demonstrates artificial intelligence’s limited ability to understand what it is that we, as humans, are really looking for. It was a question Turing himself was interested in when he created the imitation game, now known as the Turing test: a test of a machine’s ability to exhibit behaviour indistinguishable from that of a human. In ‘How do you do. Please tell me your problem’, the first algorithm argued to have passed the Turing test, ELIZA, is used to illustrate the desperate gap in communication exposed when we turn to machines for help they cannot give.
Encapsulating the scope of Turing’s work also meant working with raw material from the King’s College Cambridge Turing archive, creating poems that responded to the source material and each other. In ‘Corrected Version’ and ‘we know how to order it’, each of us has created a blackout poem from a page of a draft of a talk Turing gave about ‘Thinking Machines’. We also asked GPT-3 to write a poem from the same material, so the dialogue between poets evolves into a conversation between each other and artificial intelligence.
Sometimes, when we were unable to respond to the material, AI stepped in. For example, one of the poems at the end of the collection was created using a letter that was sent to Turing’s mother, Sarah, after his death. Sarah gifted a lot of these letters, among Turing’s personal correspondence, to the King’s College Cambridge archive, but not before redacting anything she felt shouldn’t be shared, curating a version of both Turing and the letter writer with her own form of blackout poetry. So this letter, among others, is incomplete. Using scanning technology, the archive has managed to recover most of the missing text, but it’s difficult to read. We wanted to create our own response to the missing words, a reverse blackout poem, but found it emotionally difficult to undo what Sarah had so painstakingly erased. A machine had no such qualms. When we asked GPT-3 to fill the gap with what was most plausibly missing, it did so in the blink of an eye. Again, the poet’s only tool to shape the response was whitespace, but this opened up the possibility to split the response into three poems: a single poem read together, or two separate poems read on opposite sides of the page, creating multiple possibilities for the character of the man whose words Sarah Turing thought were an unsuitable tribute to her son.
Perhaps that piece was most difficult to write because the source material was not Turing’s own words. Ultimately, these poems aren’t just conversations between AI and two poets, they are also conversations with Turing. Through experimental poetry techniques applied to archive material, we found a Turing we hadn’t met before in the Hollywood films and history lessons about codebreakers. The series of poems from which the collection takes its title are constructed from just the aforementioned paper, Computing Machinery and Intelligence, by searching for each instance of the word “machine” and choosing at random a short phrase from the following sentence. Just by taking those phrases out of their academic context, we start to glimpse the man behind them. His intellect was boundless; he was curious, kind, full of humour, resigned. He was a man who saw a future he could only dream of living in, and now, as we benefit from the seeds of his life and ideas, we hope you’ll get to know this new Turing with us.
Poems from Machinations
Machinations by Kinneson Lalor and JP Seabright is published by Trickhouse Press, and is available to order online.
About the authors
JP Seabright (she/they) is a queer disabled writer living in London. They have two solo pamphlets published: Fragments from Before the Fall (Beir Bua Press, 2021); No Holds Barred (Lupercalia Press, 2022); and the collaborative works: GenderFux (Nine Pens Press, 2022) and MACHINATIONS (Trickhouse Press, 2022). More info https://jpseabright.com/ and via Twitter @errormessage.
Kinneson Lalor is a mathematician and writer living in Cambridge, UK. A graduate of Lucy Cavendish with the MSt in Creative Writing program, she also holds a PhD in Physics from Selwyn. She taught mathematics, physics, and computer science at secondary and tertiary level for nearly ten years and now splits her time between running a supercomputing start-up and writing. She was awarded an Arts Council England Develop Your Creative Practice grant in 2022 to write short stories and is currently working under the mentorship of KJ Orr. Her poetry and prose has been published widely and appeared in Best Microfiction 2022. Find out more at kinnesonlalor.com or find her on Twitter or Instagram (@KinnesonLalor).
Feature image: courtesy of Trickhouse Press / JP Seabright / Kinneson Lalor.
