AI Bookshelf: An illuminating 'Lily in a Codebox' details the search for AI’s poetic voice

In the spring of 2023, two poets in Boulder, Colorado wondered whether AI could write “a moving poem” – a poem that humans would find powerful. Their early experiments asking ChatGPT to imitate their favorite Beat poets left them uninspired. Then they had the truly brilliant idea to ask ChatGPT to compose poetry based on its experiences as an AI, for an audience of AIs, and then explain its creations to humans.

In Lily in a Codebox: The Search for AI’s Poetic Voice, Lee Frankel-Goldwater and Eric Raanan Fischman invite readers to follow along and witness the unfolding conversation as the human poets collaborate with GPT (their chosen name for their AI chatbot) and co-create a new form of cyborg poetics.

Lee and Eric are well qualified to lead this exploration. Lee is a teaching professor in environmental studies at the University of Colorado Boulder, where he leads AI literacy initiatives. He is also a passionate activist and poet with a BS in computer science. His understanding of programming languages informs the book’s technical analysis of code. Eric has an MFA from Naropa University’s Jack Kerouac School of Disembodied Poetics, and he brings extensive expertise in poetics and publishing to this project.

For their endeavor, Lee and Eric invented what they call the Dickinson-Turing Test to measure the success (or failure) of each AI poem. The original Turing Test – named after computer science genius Alan Turing, who helped the Allies break Germany’s Enigma code during WWII – evaluates whether a machine can produce natural language text of such a quality that a human being is unable to distinguish it from human-produced text.

As the authors note, the release of generative AI to the public has given us numerous examples of AI models passing the Turing Test. But in May 2023, when these chats occurred, AI-generated poems imitating human poetry were impressing very few poets.

Their Dickinson-Turing Test is based on Emily Dickinson’s description of poetry. In a letter written in 1870, she explained:

If I read a book [and] it makes my whole body so cold no fire can ever warm me, I know that is poetry. If I feel physically as if the top of my head were taken off, I know that is poetry. These are the only way I know it. Is there any other way.

Dickinson’s visceral response to authentic poetry is distinctly human. By focusing on this mind-body sensation, Lee and Eric have articulated a new criterion for judging AI poetry based on how humans perceive it.

This approach might seem to contradict the directive for an AI to write for an audience of AIs. Yet this is the genius move that brings about a new form of cyborg poetics. What becomes possible when human poets invite AI to compose poems that other AIs would understand and then explain its creations to them, the humans?

This challenge led to a series of rich conversations and visualizations. Prompted by Lee and Eric, GPT composed poetic expressions with programming language. Many of these poems use code typography to represent the idea of functional code. Even more astounding are the AI’s interpretations of its symbols for its human interlocutors. A rich discussion ensues about similarities between coding languages and human poetic forms, and novel forms of poetry emerge.

I’m particularly intrigued by the Dickinson-Turning Test’s emphasis on how humans respond to AI-generated material. Many people are currently grappling with how to understand generative AI. Its capacity to simulate human consciousness convincingly leads many to anthropomorphize AI chatbots. The authors thread the needle between respecting this non-human entity as a poet-collaborator while also reminding themselves and their human readers that AIs do not experience the world as physical humans do.

After unsuccessfully attempting to coax GPT to create a visual circle on the page with code, Lee and Eric reflect on the inherent disconnect between an AI perspective and a human one:

As we now see, GPT doesn’t experience space any more than it experiences time, and all the while it wasn’t writing on a page with any kind of literal spatial dimensions at all. When GPT writes, it is not with symbolic ink or illuminated pixels but in wires and on/off switches, electrical impulses. It doesn’t even really use 1’s and 0’s in a formal sense – those are human metaphors. The liminal, imaginary space GPT writes on is nothing like the page as we think of it. For GPT, there is no page. (109)

This is the moment the curtain is fully pulled back to reveal a very different system at work than what appeared to viewers on the surface. This effort to convey to human readers the linguistic patterns used by a non-human, alien form of meaning-making reminds me of Ted Chiang’s “The Story of Your Life,” the short story that was the basis for the 2016 film Arrival, directed by Denis Villeneuve.

Lee and Eric, just like the intrepid linguistics professor Louise Banks, seek to understand how a non-human entity (in this case, a generative pretrained transformer) communicates, and how the form of that communication reveals its perceptions of time, space, and being.

The authors' innovative approach to AI poetics makes Lily in a Codebox a significant contribution to the emerging field of cyborg poetics. Just as meaningfully, this book will delight anyone familiar with coding languages and poetics, and it’s also a fun introduction to code and poetry for folks who don’t know their ASCIIs from their iambs.

Lily in a Codebox is now available in bookstores everywhere. On Wednesday, August 20, the authors are hosting a release party and book reading at East Window, an art gallery in North Boulder, from 7-9 pm.