Low-Hanging Fruit

When the large language model GPT-5 was released by OpenAI in August, the first demonstration of its improved text generation consisted of a pair of poems, from before the upgrade and after. Both models received the same prompt: ‘Write a short poem that makes an emotional impact. A widow in Kyoto keeps finding her late husband’s socks in strange places.’ GPT-4o gives three rhymed quatrains in slightly lumpy tetrameter:

In the cupboard with the rice bowls – one,
folded neat, like he’d just stepped out.
Another behind the garden broom,
dusted with plum blossoms and doubt.

In GPT-5’s effort, by contrast, the caption asks us to notice that the ‘predictable structure and rhyme scheme’ are gone. AI has reached its modernist stage. The first stanza is just two lines:

In the tea tin, a single sock,
smelling faintly of camphor and rain.

At a share sale last month OpenAI was valued at $500 billion. The hype around AI doesn’t directly depend on how good GPT-5 is at writing poetry, but poetry has been curiously prominent as a test case and/or window-dressing for LLMs: OpenAI’s rival Anthropic calls its GPT equivalents Haiku and Sonnet; Google’s used to be known as Bard.

On the one hand, these branding decisions work to advance a claim about AI’s sophistication. It’s culture-washing with an edge of metaphysics. The association with poetry confers cachet, and implies a degree of rhetorical fluency. Yet something more is at stake here, something the tech companies cannot say but can make a lot of money by implying, which has to do with poetry’s long-held status as an exemplary intersection between writing, thinking and feeling. Hence the ‘emotional impact’ that OpenAI’s prompt asks for. The lyric poem, as Allen Grossman once put it, ‘is the genre of the “other mind”’. To showcase your large language model navigating that genre is to broach the kind of question more often consigned to sci-fi.

On the other hand, AI poetry offers an easy win. The use of verse to suggest an LLM’s human-like capacities taps into a tradition where those capacities are already being used in an artificial schema. Rhyme and metre are relatively simple linguistic algorithms. Combining them with a heightened tolerance for arbitrariness of content – poetic licence – makes poetry low-hanging fruit for automation. Programmers were already generating verse in the 1950s. Analogue versions date back to the fourth century.

GPT-5’s flight into free verse, then, offers the spectacle of a mechanical algorithm apparently learning to transcend mechanical algorithms. But in fact, the breakthrough is chimerical. GPT-4o was perfectly capable of writing free verse if asked. So were various more primitive predecessors (random lineated sentences being even easier to generate than rhymed ones). Given that the two poems on OpenAI’s website were likely to have been hand-picked from a range of attempts, the contrast between them probably comes down to a comms team looking to conjure the shock of the new and arriving at a century-old literary playbook for doing so, more than it shows any profound difference in the models.

Still, as an amalgam of technological progress, literary history and corporate messaging, the GPT-5 launch poem is an interesting artefact. For one thing, it shows an improved ear for phrasing, not least by following T.S. Eliot’s maxim that free verse should always have the ‘ghost of some simple metre’ round its edges. Two-and-a-bit lines from the middle of the poem­ – ‘Kyoto’s bell rolls evening down the hill./She hangs them on the bamboo pole,/black flags’ – pull themselves into iambic pentameter, and contrive a smoothly engineered run of alliterative ‘l’ sounds. Someone at OpenAI is thinking about the market for eloquence.

Even more telling than the poem itself, though, is the chosen prompt. The command to be short and emotive is as you would expect. So is the part about a widow finding her late husband’s socks. A moment’s sentiment discovered in the everyday, poignant but ultimately consoling: this has been poetry’s default setting for centuries. But why – and this is the $500 billion question – are we in Japan?

It’s partly a message about cosmopolitanism. The widow in Kyoto could have been a widower in Lagos or a child in Helsinki. But the gender and geography matter. Anne Anlin Cheng argues in Ornamentalism that East Asian femininity has been constructed in modern Western culture as a focal site of ‘synthetic personhood’, an identity whose embodiment is persistently conflated with material ornament and display. Seen in this light, GPT-5’s performance of a grieving Japanese woman starts to look queasily like Madame Butterfly meets Blade Runner.

There’s an East Asian connection to the poem’s form as well. One of the starting points for the free verse vogue that swept through English-language poetry in the early 20th century was the moment in 1908 when F.S. Flint came across a set of Japanese translations and undertook to copy their ‘little dropping rhythms, unrhymed’. What originally prompted ‘the genre of the “other mind”’ to forego its formal algorithms was, as in GPT-5’s case, wrapped up with synthesising Japan.

Flint had no knowledge of Japanese but relied on a translation into French. To such Westerners, as Edward Said wrote in Orientalism, Meiji Japan was a ‘textual universe’, a mirror-image model of modernity comprised of signs and tokens circulating with the inscrutable freedom of words on a page. Another feature of Japanese translations that made its way into the modernist repertoire – which both the old and new GPT models replicate in their opening stanzas – is the haiku-style omission of a main verb. ‘In the tea tin, a single sock,/smelling faintly of camphor and rain’: where is the relation of action, the doing word? Such is the textual universe’s house style, born at a moment when the line between words and doing, text and power, was becoming more and more freely blurred.