In our first look at the relationship between poetry and technology, we test how language-generation software GPT-3’s reimagining of a Shakespeare classic holds up under literary scrutiny
GPT-3, the latest language-generation model developed by OpenAI, has made headlines for its sophisticated ability to write. A Guardian essay written by GPT-3 attracted a lot of attention, and the machine learning software has been described as “shockingly good” – though not without fault. GPT-3 is undeniably a powerful and exciting piece of language software, with 175 billion parameters. But how will it stand up under a literary critical eye?
When OpenAI made GPT-3 accessible via a private beta this summer, Gwern Branwen conducted an experiment to test the software’s capacity for creative fiction. The results are both entertaining – its failed dad jokes will appeal to those with an appetite for surreal, deadpan humour – and fascinating for an aspiring literary critic. Branwen repeats Kane Hsieh’s GPT-2 Transformer Poetry exercise, in which the previous generation of the software was prompted to rewrite a range of well-known poems, from Dr Seuss to Dante Alighieri. In the preface to Transformer Poetry, Hsieh clarifies that the goal of the book is “to share the whimsical results of mashing together two disjoint fields. This book is not meant to have any scientific or literary value”. However, literary critics have long known, to paraphrase Wimsatt and Beardsley, that the author’s intention is no standard for judging the success of a work. Which is why this essay aims to evaluate GPT-3’s attempt within a literary critical framework, not to the end of establishing value but to consider GPT-3’s poetry as a critical tool.
Branwen notes that prompting GPT-3 “just using the title/author is slightly unsatisfactory, as GPT-3 has memorized many of these famous poems and will, if you do not specify otherwise, happily complete them”. I was intrigued by the echoes with the New Critical teaching method of removing a work’s title, author and date to avoid, as IA Richards points out in Practical Criticism, stock responses and critical preconceptions that draw too heavily on a student’s memory, rather than their critical faculty. In the case of GPT-3, an additional prompt offers a solution: “Poetry classics as reimagined and rewritten by an artificial intelligence”.
For this study, I’ve turned my attention to perhaps the most famous poem in this anthology of famous poems, William Shakespeare’s Sonnet 18. The speaker’s diatribe against the summer’s day that pales in comparison to his “more lovely” subject has, in ways, become the archetypal Shakespearean sonnet: structurally, through its strict adherence to iambic pentameter and ababcdcdefefgg rhyme scheme, and comprising three quatrains including a volta (change in tone) followed by a couplet to finish. But the poem is also a testament to Shakespeare’s mastery of playing with tradition. He disrupts a value system established by the pastoral tradition by rejecting the natural idyll, instead focusing on how “nature’s changing course” is given to extremes (its “rough winds”, being too hot or too dull, etc). The speaker’s historic opening question – “Shall I compare thee to a summer’s day?” – signposts what he is not doing, as the subject of the poem remains notably absent until the volta (“But”) when the second-person pronouns reappear. The poem ends on a self-referential note: the subject’s “eternal summer shall not fade” because it is immortalised in Shakespeare’s “eternal lines”. As long people can read, they’ll read this poem about you, he says – 400 years later, he has yet to be proved wrong.
GPT-3’s reimagining of Sonnet 18 does share some qualities with the original. It is a praise of the beauty of the speaker’s “mistress”, which approximates the first half of Shakespeare’s verse and is an acceptable theme for such a poem. A generous reading might allow for a fleeting self-reference, in the poem’s explanation that “The din of merry hearts hath brought me thus | To greet thee”. In terms of coherence with Shakespeare’s first four lines, it draws on the imagery of weather, moving on to the “clouds intemperate” and lacking “beams of sun” of winter. It could be read as a eulogy, with its “shut up” eye, repetition of “fade” in the second half, and evocation of the “cold and moist” earth that “Grows as both ugly and uncourtly” without her presence.
But the poem’s most impressive feature is its application of formal elements (excepting the basic structure – the first identifying feature of a Shakespearean sonnet is its 14 lines, a quality missed or disregarded by GPT-3). Where GPT-2’s response to Sonnet 18 built on the summer imagery but not much else, relying on redundant anaphora (starting many successive lines with “And”) and earthly clichés (“the grass is green”, “the sky is blue”), GTP-3’s version draws on Shakespeare’s vocabulary, employing adjectives that do not appear in the original Sonnet 18 but are used elsewhere in Shakespeare’s oeuvre (“unbrac’d”, “saint-like”, “unsoil’d”). It also appropriates Shakespeare’s meter, even using the otherwise redundant repetition “lovely-lovely” to retain the line’s iambic pentameter.
GPT-3’s Sonnet 18 gets more unsound as it goes on, becoming nonsensically repetitive (“Neither flesh of love nor love’s herself my love”), confusing typical images (“a blush on your forehead”) and throwing caution to the wind where syntax is concerned. But, nonetheless, the poem retains the feeling of Shakespearean verse in its meter and register. Reading Shakespeare’s Sonnet 18 alongside its GPT-3 counterpart is a useful exercise in highlighting the clever features of Shakespeare’s work: its comparison of fallible nature with everlasting poetry is emphasised next to GPT-3’s images of mortality, for example. Comparing the two refreshed my appreciation of Shakespeare’s play with more intricate concepts, such as irony, that are currently beyond the remit of machine learning. GPT-3’s poem inspires questions of what it means to be Shakespeare-like without being authentically Shakespearean, an experiment that left me wondering what can be stripped away from Shakespeare’s poetry without removing its essence and which qualities lie at its heart. More generally, AI poetry brings new life to exhausted literary critical debates such as authorial intent, and asks a compelling question: is there a difference between language generation and writing?