Exploration of Co-Creation with AI



As AI has changed dramatically over the past decade to become a defining feature of the 21st century, we have to consider its interactions with humans. Since 2018, I've been experimenting with different ways that people can collaboratively create poetry with AI over the course of three projects.




2018: Poetry Exploration


Initially, this project was an experimentation with GitHub user hunkim’s tool, word-rnn-tensorflow. This recurrent neural network allowed me to generate text with the unsupervised training based on a collection of any source text.

Initial Generation
This iteration of generation an example of the result of training based on an anthology of E.E. Cummings' poetry. While it seems garbled and lacks resemblance to a poem, it has words that are somewhat reminiscent of the content of his poems. Additionally, the words make sense in relation to each other when read out as a sentence.

“the while in i it and stone all nerves thousand (with feels is kinder i yes) but candy this say and increase themselves asleep, his goldberger died five thing into the sick, (upon who man paper soon of forever and who 41 murder bird i pain is if the iful and when the other; period ness of like this more land hello and call look and slave singing”

Co-creation
I sent this poem to my friend, an avid poet. The instructions were simple: make it your own. By pulling meaning from the initial generated text through adding text and using literary devices such as enjambment through formatting the poem into lines, she was able to create a unique poem that combined both the creativity of the RNN and her own meanings.
I used this poetry as an exercise in typography.



2019: Poet Tree App


Poet Tree is an experience of collaboratively writing with AI and sharing poetry with people. I created an interactive prototype using Sketch and Principle.




Poetry Co-Creation

Based on initial explorations of poetry, I identified several key parameters of interaction that could shape how people co-create poetry. For the first parameter, users can choose a poet to generate poetry from. Based on a collection of models trained on poets like Edgar Allen Poe, the selected poet defines the vocabulary and structure of generated poetry. The next parameter is the "seed". For my model, the seed is the first set of words that are used for generation. This seed influences the content of the following words, so a poet can set the tone of poetry based on a word like "mountains" or "time". When a poet comes up with an interesting line they’d like to start a poem with, they can save it as a seed to use for later. 

Once poetry is generated with these parameters, poets can edit the text and make it their own. When sharing their work, poets can attach illustrations or images to accompany their poem.
Sharing Experience

Slam Poetry can be a powerful communal event that brings poets together through sharing poetry. This solidarity is often expressed through the norm of snapping– when audience members feel that they've been touched by poetry, they snap. Similarly, people who feel connected to a poem on Poet Tree can also snap for however many times necessary to express their appreciation.

Poets can also attach meaningful images to their poetry to complement their writing, using photography or illustration to bring further visual meaning.



2020: Immersive Installation


I aimed to create a mockup of an experience where users could be immersed in poetry that they created. Users are prompted to initially speak a single word, which would act as the "seed" in a similar way as the Poet Tree App. Line by line, AI-generated poetry is projected onto visuals to immerse and inspire people to co-create with AI and famous poets.


2021: Poet Tree Website


Using GPT-3, a powerful language processing tool by OpenAI, I’m currently redesigning and developing an app where poets, rap artists, and creative writers can engage with AI in their creative process.