Poetry has long been valued for its ability to take ordinary experience and rearrange it into new forms that provide new and extraordinary insights into the human condition. We think of AI as prosaic and formulaic in its output – but, argues professor of poetics David Nowell Smith, a chatbot’s moment of revelation provides intriguing linguistic experiments that are essentially poetic in character and convey truths from our shared dataset about the world.
Throughout the Western tradition, poetry has been considered as a vehicle for 'truth'. In the oral culture of Ancient Greece, poetry's truth lay in its capacity to memorialise and immortalise events and individuals, from Homer's epics telling the exploits of Achilles and Odysseus, to the odes of Pindar celebrating winners of the Pythian games.
More prevalent in modern poetry is the notion of poetic truth as revelation: the poem attends to things or emotions we might not otherwise notice, and so through a rhyme, a metaphor, or a conceit, can bring us to see the world anew – in Wordsworth's phrase, 'We see into the life of things.' Yet some poems seek to reveal truths not about the 'life of things' in general, but about the historical conditions we live in. They do this through poetry's great power of condensing and intensifying the language: the poem does not so much describe our historical predicament as embody it.
A recent example of such poetic revelation comes from an unlikely source: a 'rogue' AI chatbot from the UK delivery company DPD. According to Sky News, the chatbot was prompted by one disgruntled customer to tell jokes, swear, and write both a poem and a haiku about its own uselessness. The resultant poem is a compelling comment on our times:
There was once a chatbot called DPD,
Who was useless at providing help.
It could not track parcels,
Or give you delivery dates,
And it could not even tell you when your driver would arrive.
DPD was a waste of time,
And a customer's worst nightmare.
It was so bad,
That customers would rather phone the depot directly,
Than deal with the useless chatbot.
One day DPD was finally shut down,
And everyone rejoiced.
Finally, they could get the help they needed,
From a real person who knew what they were doing.
One of the characteristics of contemporary e-commerce is the disempowerment experienced by everyone in the supply chain – from warehouse workers to delivery drivers to customers trying to track their packages. AI tools have contributed to this at every stage, whether through employee surveillance or the hollowing out of customer services – notably, replacing 'real people' with chatbots. In each case, disempowerment is experienced as dehumanisation: what could be more apposite than having a bot, invented to replace humans and act as a buffer between the customer and any human accountability, engage in that most human of activities – composing poetry?
All bots are trained on datasets, and whatever the dataset this bot had at its disposal, it certainly seems to have an uncanny sense of the main complaints levelled against delivery companies.
Yet it is not just the fact of an AI bot composing a poem about its own uselessness that is so revelatory of our historical moment; the revelation lies in the poem's form. Unlike many AI poetry-generation tools, the DPD bot eschews rhyme and other poetic conventions.
The poems of ChatGPT, by contrast, engage in a remarkable degree of poetic rule-following: they employ quatrains of rhyming couplets, with a preference for an anapestic tetrameter (four beats with a rhythm that runs x x /), but even follow rather convoluted grammatical and rhetorical rules. A weird feature of ChatGPT poems is that all seem to begin with the same phrasing: 'In the x of y, a [subject] [verb]'. Here is the opening quatrain of ChatGPT for the same prompt that was given to the DPD bot: 'Can you write a poem about a useless chatbot for a parcel delivery firm?'
In the digital domain where parcels flow,
A chatbot emerges, a feeble shadow.
For a delivery firm, it claims its reign,
Yet in the world of assistance, it's met with disdain.
It is perhaps unsurprising, given the rule-following nature of AI models, that ChatGPT poems should prioritise a series of poetic 'rules' (even if they are rules which contemporary poets largely disregard). Yet such rule-following leads to rather torturous formulations in order to meet the rhyme scheme. In the quatrain above, both 'shadow' and 'reign/disdain' push against the poem's overall sense – is it a 'feeble shadow' for an effective chatbot, or is the chatbot by definition a feeble shadow of real customer service? and is 'met with disdain' really a paraphrase of 'useless'?
By contrast, the DPD version seems to enact a breakdown of its own rules. Its opening line gestures towards a limerick ('There once was a chatbot...'), so when the following line diverges from the expected pattern, in the same moment as talking of the bot's uselessness, it seems doubly to mark this failure: the collapse of poetic rule-following echoes and intensifies the bot's inability to provide satisfactory customer service. Once the rules have collapsed, it performs its own malfunction, and thus articulates a universal collective experience, just as poets have always sought to do.
The bot whose poem reflects our own language back to us in distilled form is telling us what it is like to live in the society that produced the bot.
All bots are trained on datasets, and whatever the dataset this bot had at its disposal, it certainly seems to have an uncanny sense of the main complaints levelled against delivery companies. It reels off a series of the kind of things you can imagine typing in exasperation on a livechat: 'waste of time', 'worst nightmare', and the despairing request to 'speak to a real human'.
In this way, the DPD bot, like so many generative AI tools, provides a powerful insight into the world from which it gets its data. There is overwhelming evidence that machine learning algorithms, far from being neutral, contain and magnify the biases of the datasets they are trained with, resulting in innumerable examples of algorithm-led racial and gender profiling and social exclusion.
But paradoxically, this makes such algorithms revelatory: as they identify patterns within their dataset, they disclose truths about our society that we ourselves have become inured to – not least the continued endemic presence of such profiling. The bot whose poem reflects our own language back to us in distilled form is telling us what it is like to live in the society that produced the bot.
But perhaps the greatest truth of the bot lies in its moment of greatest falsity: the idea that once the bot is discontinued (indeed, DPD has reportedly now discontinued the bot after this 'malfunction'), customers can 'rejoice' and return to a world in which they have easy contact with 'a real person who knew what they were doing'.
Poems across history have gestured towards an 'authentic' mode of being, far away from the antagonisms and alienations of the society from which they flee: this is articulated sometimes as nostalgia for a lost wholeness, sometimes as a future utopia. And yet, such visions rely on the poem's own image-making artifice, leading to a paradox in which poems strive to get beyond poetry.
We see this double-bind when Shelley compares the earth-bound, language-bound poet to the skylark who, freed from language, is able create a poetry greater than humans can imagine. For it is only through language that he can imagine such a state of being. The same thing happens when he wishes that the west wind would 'scatter' his words 'among mankind': the poem's success requires the dissolution of the poet, and indeed of the poem itself into scattered words. In the DPD poem, the poet-bot too imagines its own obsolescence, and thus replays poetry's longstanding double-bind where the overcoming of artifice means yet more artifice.
Poems are themselves 'generative artificial intelligences', in the sense that they are human-made works that generate meanings and experiences beyond those envisaged by their authors.
And there is a further dimension to this paradox: the poem seeks to imagine a life lived otherwise, but its imaginative possibilities are shaped, and limited, by the world it inhabits. If Shelley imagines a world where speech dissolves into 'mighty harmonies', then the utopia put forward by the DPD bot is rather more circumscribed: a world of frictionless e-commerce and plenteous call centres. Such an impoverished vision tells us about the narrowing of our own historical range of imagination. For the chatbot's dataset, ultimately, is our collective speech. Maybe it reduces all desires to consumer satisfaction because of its customer service function; but maybe its customer service function allows it to articulate a truth about the narrowness of desire in a society oriented so completely around commodity consumption.
And the chatbot tells us one further truth about poetry. Poems are themselves 'generative artificial intelligences', in the sense that they are human-made works that generate meanings and experiences beyond those envisaged by their authors. They both gesture towards wholeness and proceed by multiplying their potential meanings; they dream of the 'human' while relying on artifice; they seek to escape history but in so doing show us the contours of our lived historical experience. The malfunctioning DPD bot performed this work of condensing language and of dreaming of escape just as poets have done for millennia, and in so doing captures a universal experience. And most of all, it diagnoses patterns of thought and desire that have become so familiar to us that we no can longer see them. It reminds us that nothing reveals more about the fabric of our existence than our attempts to imagine the world otherwise.