Machine Learning Meets Verse at the ITP 30 Show

The ITP 30 Show gathers creative technologists from NYU's Interactive Telecommunications Program and lets them present projects that blur the line between code and culture. Among the most striking entries are the works that use machine learning to compose poetry, training neural networks on everything from classical sonnets to contemporary spoken word. The participants approached the question of algorithmic verse as designers, engineers and poets, treating each generated line as a collaboration between human intention and statistical pattern recognition.

For Australian audiences, this kind of work resonates strongly with a national creative scene that stretches from the laneway galleries of Melbourne to the festival circuits of Adelaide. Communities in Sydney, Brisbane, Perth and Hobart have been quietly nurturing their own pockets of computational creativity, often borrowing from the same open-source toolkits that the ITP cohort used. The conversations happening at the show echo debates playing out in Australian cafés, classrooms and studios about what authorship means when a model can produce a passable stanza in seconds.

Choosing the Corpus and Training the Model

Several participants began by curating the textual diet of their models, feeding them carefully selected collections of verse. Some sourced public-domain poetry from archives of Romantic and Modernist works, while others scraped contemporary blogs and zines to capture looser, more conversational rhythms. A common technique was to fine-tune a base language model on a small, stylistically coherent set of poems, then sample from the resulting network at low temperatures to keep the language recognisably poetic rather than drifting into prose.

This attention to corpus echoes the careful selection processes Australian poets use when building their own collections, whether drawing on the work of the Bundjalung and Noongar oral traditions or the experimental verse published in small press journals out of Fitzroy and Newtown. The participants treated training data as a kind of literary terroir, recognising that a model trained only on haiku would produce a very different flavour of output than one trained on sprawling free verse or hip-hop lyrics.

Prompt Engineering as a New Poetic Craft

Once the models were ready, contributors spent hours crafting prompts that could nudge the systems toward specific moods, voices and structures. Some composed template strings that forced the network into strict forms, asking for fourteen lines of iambic pentameter or a sequence of tanka about a single object. Others experimented with negative prompts, instructing the model to avoid clichés, sentimentality or generic imagery. The act of writing prompts became its own form of authorship, sitting somewhere between programming and lyric composition.

The same instinct shows up in the way Australian digital artists describe their process in studios from Collingwood to Fremantle. Practitioners there often speak of working with generative tools the way a barista pulls a flat white, adjusting parameters until the output feels right. Participants in the show adopted a similar vocabulary, describing their prompts as recipes and their model settings as grind size, tweaking until each generated poem achieved the right body and crema.

Hybrid Systems and Human Curation

Almost none of the projects presented at the show claimed that the machine alone produced the final poem. Instead, contributors built hybrid pipelines where the model generated dozens or hundreds of candidates, and a human editor selected, revised and reordered the lines. Some used the algorithm to draft an ending they could not find themselves; others used it to break a creative block by surfacing unexpected metaphors. The resulting works were deliberately co-authored, and the participants were transparent about which lines came from the model and which were rewritten by hand.

This collaborative sensibility fits comfortably with Australian creative practice, where collective studios and shared authorship are common in the indie music scenes of Brunswick and the design collectives of Surry Hills. The ITP participants treated their models less as autonomous poets and more as unusual collaborators, much like the way a producer might work with a vocalist who speaks in a language they only half understand.

Sound, Image and the Interactive Page

Several poetry projects were not just texts on a screen. One team paired generated verses with generative illustrations, building an illustrated catalog that responded to the emotional register of each line. Another turned a poem into a weather-based experience, displaying stanzas that shifted with the actual forecast for the reader's location. A third project produced an interactive archive where visitors could adjust sliders to steer the model toward different rhyme schemes or lexical registers, producing a custom poem in real time.

These interfaces suggest the same appetite for playful, place-aware technology that shows up in Australian experiments at places like MONA in Hobart, where art and code frequently share the same room. The ITP cohort treated the poem as a living object, something that could be performed, illustrated or heard, rather than a static block of text.

Copyright, Consent and the Question of Style

Working with machine-generated verse quickly raised legal and ethical questions that participants addressed openly. Several explicitly avoided training on copyrighted contemporary collections, preferring public-domain or openly licensed material. Others documented their sources and released their own corpora under Creative Commons licences, modelling a transparent approach that would satisfy the disclosure expectations found in Australia's privacy and copyright frameworks. A few projects explored the question of stylistic imitation, asking whether a model trained on a particular poet's body of work was paying tribute or borrowing without consent. Visitors who want to continue these conversations are warmly invited to reach out through the project's contact page.

The ITP 30 Show does not pretend these questions are settled, and visitors are encouraged to bring their own perspective to the conversation. Full credits and biographical notes for every participant are gathered on the authors page, where the names behind each model, poem and interface can be explored in detail, and where the human collaborators who shaped each generated line can finally be acknowledged by name.