Let’s start with two poems. Glance over the first one and consider how it makes you feel. Ponder the author’s message: what might have inspired them to write this? For best results, jot down the ideas and emotions the poem evokes for comparison later.
After absorbing this poem, repeat the exercise with the next one. What visual, mental imagery is stimulated by this poem? Do you feel a personal connection to the text? Take notes.
Perhaps surprisingly, both of these poems were authored by a bot, and they are meant to communicate a very specific image. Bei Liu, Jianlong Fu, Makoto P. Kato, and Masatoshi Yoshikawa created a bot which was trained to generate poetry from a singular image. Using technology analogous to that which automatically suggests descriptive captions for the hearing and visually impaired, the team took it one step farther and used a poetry database to teach the bot poetic language and symbols. Their paper “Beyond Narrative Description: Generating Poetry from Images by Multi-Adversarial Training” explains the process whereby they created different models, compared them, and selected the most effective model for testing (in their case, finding out if people could tell which poems were AI-generated). They published eight of the chosen image-poem pairs, and from those, our experiment was born.
Of these pairs, I chose three with disparate images to investigate. I made one survey of all three images and asked 55 people to characterize the response the image evoked. With 55 descriptive lists of ideas, emotions, and visualizations, I could see how real people reacted to the same images the bot was given. Then, I took the correlating poems and repeated the survey — this time asking 55 new participants to react to the poem, in addition to asking a few more questions like “Did you find this poem meaningful?”
Then it got exciting. I took the set of 110 responses and for each image and poem organized it so that every word correlating to image 1, for example, was pooled together. Then, I used these response pools to produce individual word maps for side-by-side comparison. Through this process, I could determine if the bot’s poetry was successful at communicating the mood of the images — and the results were fascinating.
Here, I have each image and generated poem adjacent and their word maps underneath. Take out your notes and see how you compare. First up is the most promising image-poem pair. As you can see, the bot was outstanding at evoking sentiments from the poem that reflected those from the image.
These results imply the possibility that artificial intelligence could soon be recognized as genuinely “creative.” The poem embodies the image very well, communicating as good poetry should, and it speaks to its audience: 52.8% of respondents claim that “this poem gave me both a feeling and mental picture,” while “43.4% agree that “this poem gave [them] a vivid mental picture” – that’s a more than 95% success rate. Beyond that, 81% of respondents found the poem meaningful, which speaks loudly given poetry’s subjectivity. Interestingly, our poet’s amazing success begins and ends here. As we see in the following examples, the bot can replicate the meaning but never the message as it did above.
Glancing through these word maps, the contrast is immediately stark: almost none of the responses overlap, and one that does carries no real meaning but merely reflects the wording of the prompt: “image.” That aside, the poem fails to convey its message. “Hidden,” “red,” “culture,” “beauty,” “mystery” — none of these ideas arose in the minds of the readers of the poem. Still, it wasn’t a complete failure. Although it didn’t communicate the image well, it did communicate something. 87% of respondents agreed that the poem gave them a “vivid mental picture,” “a feeling,” or both, and 63% agreed that the poem was meaningful. And to be fair, there are some similarities: “lost” and “hidden,” “mystery” and “confusion,” “girl” and “woman” are definitely related pairs. With that in mind, this pair takes second place, boasting the ability to communicate some kind of meaning and a few synonymous sentiments.
Now, for my favorite comparison. In what appears to be everyone’s least favorite image and least favorite poem, we have a contrast that might not be able to overcome its differences. With essences decidedly at odds and less meaning than ever before, this image-poem pair works to fight the notion that AI ever achieved the ability to communicate through poetry.
As you can see, the very marrow of the responses differ in nature: “death,” “decay,” “dead,” “gross,” and “sad” have nothing to do with “blue,” “sky,” “feeling,” “bright,” or “peaceful.” Surely, “crab” and “beach” correlate with “water” and “ocean,” but there is clearly no visible water in the image, and an overwhelming number of responses to the image maintained a focus on death. While 85% of participants found either a feeling or mental image from this poem, a record low of just 50.9% found it meaningful — just one more participant than those who found no meaning at all. Coming in last place, this duo threatens the bot’s ability to claim a seat at the poet’s table.
But at the end of the day, who gets to make that call? Who can say, to those who found all three poems meaningful, that because no one realized this poem should have focused on death, it doesn’t count? And anyway, who even decided that a poet has to communicate the popular sentiment? If all poets did that, creativity would cease to exist. Who is the gatekeeper to the world of art?
The short answer? You. And me, and every individual person on their own. At the end of the day, this bot already created these poems. They exist, and it’s up to us to decide if they are valuable and what they mean. To do this, definitions matter. For some, art is simply anything created with the point of bringing beauty into the world, while others think it’s a way of creatively expressing oneself. If you are rooted in the school of thought which focuses on intentionality, the belief that it is the author’s intentional message that gives life to art, these poems are no more meaningful than a label of nutrition facts. But, if you follow the reader-response criticism which focuses on an audiences’ reaction to a piece, you might find yourself supporting an AI-generated poetry publication. And if you’re in between? Well, welcome to the club.