AI as Poetic Technology: Sasha Stiles' Living Poem

Sasha Stiles turned GPT-2 experiments into a self-writing poem at a Museum of Modern Art installation—and a new way to think about text-generating AI optimization. Poetry and artificial intelligence can appear as opposites—one deeply human; the other cold and mechanical. Sasha Stiles sees them as expressions of the same impulse. Poetry, the Kalmyk-American poet argues, is “one of our most ancient and enduring technologies,” a system of meter and rhyme invented to store vital information. She views AI as its natural heir.

Stiles’s path to AI began with literature, not code. But science was never distant: Her parents are documentary filmmakers who worked with Carl Sagan on the original Cosmos series, and she grew up traveling with them as they interviewed scientists and philosophers. She came of age with the Internet and sensed how it shaped the way she thought and wrote. When she encountered the technology underpinning modern AI in 2019, she didn’t want to just write about it—she wanted to write with it.

How did you end up creating art at the intersection of poetry and AI? In 2017 I read about the transformer-based architectures that drive natural language processing, and something clicked. That’s the moment I realized I didn’t want to just write around the idea—I wanted to understand what it felt like to write using those models, to use AI as a tool, as a collaborator and as a medium. I started doing research and looking into who was using these early versions, who was doing interesting things with language. This was long before ChatGPT. I was reading the work of people like Ross Goodwin and Allison Parrish, who were creating poetry bots on Twitter. And I tried to figure out how someone with a humanities background could take baby steps into that world.

What did those early sessions look like? I started working with GPT-2 in 2019, taking lines of my own poetry and feeding them in to see what would happen if I asked a language model to take an idea I’d had and run with it. One of the first poems I wrote came from inputting the line “Are you ready for the future?” over and over again into the same system, tweaking the parameters to see how the output would change. It wasn’t intended to be a poem—just research. But I found it really interesting and ended up curating 30 of those hundreds of outputs into a little poetry cycle. The results ranged from very sublime and beautiful to very misogynistic or pornographic—really looking at the spectrum of what you were able to output at that moment.

How did you go from feeding lines into a generic model to building something trained on your own voice? I took a manuscript I had pretty much done at that point—200 pages of poems—and put it all into a dataset to create a fine-tuned version of GPT-2. So I had a system that actually had knowledge of my own writing—not just the canonical poetry already in the archives but a sense of my style, my vernacular, my thematic areas of fascination. I could use lines of my own poetry as inputs, knowing the system had a sense of how I had already written that poem. That process eventually led to A Living Poem at the Museum of Modern Art in New York City.

How does that piece work? I think of A Living Poem as a living language system—an hour-long script for an unfixed, ever evolving poem in which code sketches, datasets, prompt architectures and human influence converge to perform real-time loops of verse, visuals and voice. It’s essentially an environment in which I can think about language and let language think about itself while making that process tangible. I’ve long been drawn to metapoetics and to generative or automatic writing in its many forms. One of my earliest encounters with computational text art was The House of Dust (1967), by Fluxus poet Alison Knowles—an early computer-generated poem. And contemporary language artists such as Holzer, Ruscha and Kruger have been formative for me. A Living Poem is rooted in this lineage and in the 20th-century technological and cultural conditions—mass media, broadcast culture, industrial printing, early computing—that shaped modern text art, which in turn shaped me. At the same time, it’s a place where I can experiment with new modes of expression emerging from 21st-century technocultural conditions: language as a living, generative field where meaning is made at speed and scale through recursion, probability, multiplicity, networked imagination.

You’ve described poetry itself as a technology. What do you mean by that? A lot of people think poetry and technology are antithetical, but I find them resonant. Poetry isn’t just an art form or decorative language. Humans invented poetic language before we had written alphabets because we needed a means to store information, preserve it and transmit it through generations. We invented meter and rhythm and rhyme so we could remember really important human data. Poetry is one of our most ancient and enduring technologies—a very primal data storage system.

Does that change how you view AI? Looking at AI through the lens of poetry is a way of saying there’s something very human about the fundamental impulses behind technologies like AI. I think about poetry as our most ancient hybrid intelligence, a way of interlinking algorithm and emotion very much in the way our new technologies are. If we can acknowledge that these technologies enabled self-awareness, consciousness and our ability to articulate inner worlds—maybe that’s useful to conversations around artificial intelligence now. Maybe these tools can take us into new territories of consciousness, just as poetry has enabled us to do for thousands of years.