how AI poetry betrays our desire for human connection
- Written by Judith Bishop, Tracey Banivanua Mar Fellow, La Trobe University
Could I but ride indefiniteAs doth the meadow-beeAnd visit only where I liked,And no man visit me,
I’d soar above the verdant hills,In search of hidden glades,Where sunlight filters through the trees,And dances in the shades.
The breeze would be my constant friend,The sky my boundless roof,And in the hush of nature’s breath,I’d find my living proof.
For in the silence of the wild,My soul would find its peace,A gentle, endless wandering,Where all my cares would cease,In nature’s vast embrace,My heart’s true, quiet lease.
Do you like Emily Dickinson? Her poetry is beautiful, isn’t it? Original, evocative, vivid. The rhymes are subtle. The lines flow softly from one image to the next.
If you thought these stanzas were written by Dickinson, you are not alone. All but one of the eight students who participated in our pilot study on responses to AI-generated poetry thought so too.
Yet only the first four lines are Dickinson’s. The remaining lines were generated by OpenAI’s GPT4o. You can read what Dickinson actually wrote here.
Emotional reactions
Our study set out to explore students’ responses to AI-generated poems. We especially wanted to know if AI poetry could produce the same feelings of emotional connection and growth that have long been associated with reading literature.
Researchers have found that both lay and professional readers of literature prefer AI-generated texts to human ones, and the majority of readers are unable to reliably distinguish between them.
A recent New York Times quiz confirmed these findings: more than 50% of the 86,000+ quiz-takers preferred the AI texts.
And yet, as the newspaper’s technology columnist Kevin Roose reported in a recent podcast, many quiz-takers were upset when informed of this. So were the students in our experiment.
What do these emotional reactions suggest is at stake when we read AI-generated writing? The results of our study may offer some clues.
Dickinson’s poetry is beloved for its emotional resonance; she has been described as “someone with whom people have a very personal relationship”.
Would AI versions of her poetry have the same effect?
To investigate this question, we used GPT4o to produce new versions of Dickinson poems that are available in the public domain. Prompting the model to reuse the first stanza, we asked it to complete each poem three times, closely mimicking Dickinson’s signature tone and style.
Students were invited to read and comment on one of two original poems and the three AI renderings. In surveys and interviews undertaken immediately after, we asked each student about their impressions and their preferences.
Students knew that only one of the poems would be “real”, but we revealed which one it was only at the end of the interviews.
An unpredictable arc
The AI-generated versions of the meadow-bee poem consistently offered a fluent and logical story of the bee as the image of a free and authentic self, released into nature. Dickinson’s original poem has a far more unpredictable arc. It gives us a picture of a fractured and ironic psychological state.
In Dickinson’s poem, the flitting bee is free to follow its “indefinite” whims and “flirt all day with buttercups”. The speaker likewise desires to
marry whom I may,And dwell a little everywhere.
But this desire quickly unravels into darker territory. She declares that she would rather flee from moral authorities (“Or better, run away / With no police to follow”) than enjoy the limited freedoms of conventional courtship, even to the point of desperate measures:
Till I should jump peninsulasTo get away from you.
The final stanza performs a brilliant and surprising return to the opening motif. As if waking from a reverie, Dickinson restates the bee theme:
I said, but just to be a beeUpon a raft of air,And row in nowhere all day long.
The poem’s wild ride isn’t done with us yet. In the final two lines, the speaker and the reader realise that her psychological situation is in fact a prison:
What liberty! So captives deemWho tight in dungeons are.
Disappointment and betrayal
Read alongside the soothing lines of the AI version, Dickinson’s original may sound fervent, harsh or depressing. What would readers think of the differences between them?
Neither we nor the students were prepared for what we found. Not only were seven out of eight students unable to tell the real Dickinson poems apart from the AI-generated versions, but they confidently rejected the original poem as “definitely” AI-generated.
What’s more, they described intensely positive feelings in response to the AI poems.
Students were suspicious of Dickinson’s liberal use of em-dashes (a hallmark of GPT writing). They dismissed her unusual or disjunctive imagery as “stilted” and “distinctly unevocative and jarring”.
In their view, Dickinson’s poem lacked the flowing, immersive, emotionally engaging and enjoyable qualities they associated with human-written poetry.
Conversely, they responded strongly to the emotional and aesthetic qualities of the AI poems, describing them as “very soothing”, “flowing”, “serene”, “calming” and “relatable”. The AI poems moved them, connected with their memories, and motivated them to experience nature.
One student found the AI-generated nature imagery “very peaceful […] as I feel when I have been out in nature”. Several students intuited the presence of lived emotion in the AI poems: “when you are reading the imagery, you feel like there is a feeling to it, I suppose”.
The same students described experiencing disappointment, alarm, fear and even betrayal when we revealed the original poem.
Moreover, knowing that a poem was written by AI produced an immediate emotional distance from the text. One student explained,“you’re not going to have that connection […] you’re seeking that connection with the human being”.
Another student said that the poem they thought was AI-generated felt “like a computer in my heart”.
A shared connection
At a time when AI-generated texts can pass for human writing, it feels critical to understand why the connection with human experience and emotions through literature continues to matter so much to us – and what we can do to preserve that connection.
We know that reading literature can help us to work though difficult emotions. The speakers of Dickinson’s poems are often grappling with intense states of mind and divided or thwarted feelings. As such, they can accompany us in recognising and understanding our own emotions.
Literature can even shift us into new ways of thinking and feeling through the internal dialogue we enter into with a text and its characters. As literature scholar Annette Federico points out, “literary reading is still fundamentally relational, or intersubjective: a self-to-other, self-to-world, and even self-to-self experience”.
Crucially, it can help us feel a little less alone. As novelist James Baldwin said:
You think your pain and your heartbreak are unprecedented in the history of the world, but then you read. It was books that taught me that the things that tormented me most were the very things that connected me with all the people who were alive, who had ever been alive.
Our research suggests that literature’s recuperative effect is amplified when reading together. When we share the reading experience, literature becomes “a springboard for broader reflections on life and personal experience”.
Valuing the surprise and idiosyncrasy of human writing, with its genuine interiority, may be the key here. Generative AI now offers production-line quality in writing across a wide range of genres. But the price may be losing our connection to each other’s human experience and emotion through our words.
We will give the final word to one of our students:
poetry is, at its core, an expressive form of art that relies on emotion and […] human experiences and the imperfections that come with being human […] I just don’t feel that great about technology trying to replicate that.
Authors: Judith Bishop, Tracey Banivanua Mar Fellow, La Trobe University
.jpg){kind=link}
