In 2015, when neural networks were still little more than laboratory curiosities, Merzmensch began compiling evidence. Not with scissors, as Tzara once did, but through Google’s Deep Dream and the early GANs, where memory itself could be trained, distorted, and replayed. These first experiments were less inventions than continuations: traces in a lineage that runs from Dada’s cut-ups to Surrealist Cadavre Exquis, from Barthes’s intertextual swarm to the post-anthropocentric murmurs of today. His case is clear, AI is not an interruption of poetry but another witness, another participant in its ongoing scene of collaboration.
What results are not poems in the conventional sense but documents of encounter. The bulk of machine output is dismissed, routine, formulaic, unconvincing. Yet anomalies remain: stutters of Beckett, absurdities reminiscent of Lynch, recursive loops worthy of Oulipo. These fragments are preserved, not corrected, and entered into the record as they stand. They operate as exhibits: objects found in the wild of computation, artifacts that testify to an emergent form of authorship neither strictly human nor wholly machinic.
To approach these works is to enter an archive of residues and misfires, where poetry is less a finished product than a forensic trace. In Merzmensch’s practice, the machine is both material and witness, reflecting our culture only to warp it, exposing seams of bias, repetition, and invention. What remains are poems that fail, poems that astonish, records of a dialogue in progress. They belong not to us, not to them, but to the evidence itself.
You’ve worked with AI as a creative partner for years. If the goal were not visual art but poetry, what would you keep from your own collaborative methods, and what would you discard?
Indeed, even if my first encounters with Generative AI can be dated back to 2015, as I experimented with Google Deep Dreams (https://medium.com/merzazine/deep-dream-comes-true-eafb97df6cc5?sk=6d50ebb59584b487183385009ba50f54) and later Generative Adversarial Networks (https://medium.com/data-science/visual-ai-trained-on-my-memories-a54c55967c28?sk=3e37e2716ccee2101ba07025e5bcfd60), my primary focus at the beginning was on poetry and textual work.
 |
| Self-portrait as Merzmensch, 2015 (Google Deep Dream) |
I’ve recognized in this kind of collaboration with AI a direct continuation of the Avant-Garde's experiments. Let’s think back about Tristan Tzara and his Manifesto (https://medium.com/merzazine/lakrobuchi-part-2-four-horned-unicorns-12865a0b4590?sk=99ee4daa9f9dc92a7fd7f59e01b4a373), in which the poem is generated randomly (by accidental combinations of cut-out words), but still: “The poem will resemble you” (Tzara).
Or Surrealists with their Cadavre Exquis (https://medium.com/merzazine/today-is-fridai-time-travels-literary-games-and-playing-exquisite-corpse-with-artificial-eaf7c43d9095), another collaborative creation with controlled actions - yet all participants are “prompting” their counterparts.
For me, it was a fascinating process to run it with a machine. How does AI interpret human culture? How does our civilization look like through the prism of neural networks? These questions motivated me to incorporate different AI approaches in my creative journey.
Yet my second approach went even further: to stop collaborating and to let the machine write by itself. A controversial take, which is surely reflected in the weltangst by many creatives: the machine becomes creative by itself. For me, this fear remains unfounded. There are a lot of writers out there. Do they really see each other as competitors? Are there artisans or rather artists? Human culture is based on intertextuality, or as Roland Barthes wrote:
“a text consists of multiple writings, issuing from several cultures and entering into dialogue with each other, into parody, into contestation;” (“Death of the Author”, Aspen 5+6, https://ubu.punctumbooks.com/aspen/aspen5and6/threeEssays.html#barthes). And if the authors observe a machine entering their scene, why not embrace the new narratives? Why not appreciate new ways of finding meanings beyond words?
The answer is simple: many are afraid of being cannibalized. But even if a machine generates millions of content pieces, it does not diminish a single line of human writing. Or put it another way: Arthur Rimbaud created around 100 poems in his life. His contemporary, Victor Hugo, wrote ca. 3000 poems. Had Hugo cannibalized Rimbaud?
Imagine now a machine creating poems in vast, almost inflationary quantities and publishing them as on-demand books on Amazon. Would these writers—because behind every AI there is still a human—replace other poets if their/machine’s work were just mediocre? Who would buy and read such poetry, who would adore and celebrate it? And if, on the contrary, the machine’s verses were genuinely inspiring, moving, even mesmerising—would that be harmful to other, human poets?
After all, it isn’t about quantity - it’s about quality, and quality in terms of poetry is a very subjective matter. AI cannot feel emotions, yet its creations can stir them in us. And I know many poems of machine origins, which touch and move me deeply.
If a machine carries the biases of its training data, can we ever expect it to produce poetry that challenges those biases—or must the human always act as the critical counterweight?
The thing is, machines are as biased as our human civilization, which means they're very biased. And if we don’t like our reflection, the last thing to blame is the mirror. We probably will never be able to liberate machines from biases, like we will never liberate ourselves. Define “bias”. Who might take that holy duty to decide what’s right and what’s wrong opinionatively? We see right now in different countries around the world how diverse the ethical concepts are, and how dogmatically they (also: we) protect our values, and offend the foreign. No, we, humans, are biased. We should deal with the plank in our eye first, before claiming about speck of sawdust in the eye of the machine.
Yes, thanks to AI, we can visualise such biases, expose them, and play with them. Humans should be a critical counterweight to humans and machines in the same intensity. So seen, machine-produced poetry challenges our own biases, and we should embrace it and scrutinize ourselves, as we do with machines. The second part is surely the easier one, so people become neo-Luddites in their narcissistic, anthropocentric self-righteousness. Saving humanity from being endangered by machines is a comforting narrative to dress up our dogmatic beliefs into a humanist messianism.
We, humans, know it pretty well, of what we are capable (in good and bad ways). And we are always up to projecting our own maleficent potential on the machines. Yet a post-anthropocentric age is coming.
In your experience, does the machine contribute more as a generator of raw material or as a structural thinker—a kind of co-architect for meaning?
Both, because I have several approaches to this. But I tend instead to let AI write by itself.
For example, I train (and not just “finetune”) language models on my own poems, mixing them with public domain books and sources. Then the language model generates a myriad of texts in various genres, with various topics, stories, and poems. This is how my series reMERZ (https://medium.com/merzazine/remerz-7920de573f70) works. For me, this is initially raw material. Ready-made. Objet trouvé. And I deal with these texts accordingly: by reading them, finding gems, and presenting them to humans. Even if it’s a “raw material”, I try to keep it as it is - but not always. We’ve seen that even the most famous ready-made by Duchamp is modified by “his” signature “R.Mutt”.
The majority of texts (90%) lack convincing poetic power. Yet around 7% are somehow interesting, so I select them for further use in my works, stories, screenplays. But the rest, 3% - oh yes! Here we can find unique irony of the machine, Beckett-esque communication failings, Lynch-esque twists, Oulipo-alike experiments, and even elements only typical for machines.
These poems I use 1:1 as the original, without intervention.
Here is an ambiguous poem, my favourite, written entirely by the machine:
 |
| “Crazy”, from: EtherPoems 2, 2021 |
Does the psychiatrist call the narrator “crazy” just to encourage him (“Come on, are you crazy? Keep writing, don’t stop it, throw your doubts away”). Or is it already a diagnosis (surely not really professional of a psychiatrist).
In another poem, it created an enigmatic world, which are slipping away from our understanding, and which is still stable and self-aware behind the gate to alternative dimension:
https://vimeo.com/572999712?fl=pl&fe=ti
“Emissary”, from: EtherPoems 2, 2021
.Post-athropocentric creatives among us won’t immediately call this apposition of concepts “BS of a stochastic parrot”, but rather will join in, co-playing the Demiurg of unknown reality, interpreting and completing this strangeness with our imagination.
This poem, for example, called “Food Stamps” seems to be an agglomeration of concepts and meaningless topics, but its absurdity is especially now very urgent and actual: it’s about decay of society and government becoming dystopian one:
Poetry is about feelings (heart) and understandings (mind), in different dimensions and intensities. Yet the perception counts. There is no poetry without readers - be it even the authors reading their own texts. Kurt Schwitters wrote his Ursonate in solitary existence outside civilization - on the Norwegian Island Hjertøya. He was the only reader of his poetry during this period. Yet he couldn’t stop writing. That’s real creativity for me - this urge to express yourself.
In my other project, KI-MERZ-AI, I trained a language model on Schwitters’ MERZ-Magazines and continued his work into the AI-driven epoch. Here as well, I hadn’t really intervened. I presented the works as they are - German literary critic Hannes Bajohr described my project as “Affirmative Kontrollabgabe” (“Affirmative Ceding of Control”) - and this definition precisely sums up my project.
(Some Images from KI-MERZ-AI)
AI tends to reproduce learned patterns. Poetry often thrives on breaking them. How might this tension be navigated in a way that produces something neither human nor machine could make alone?
It depends. Here you have to stop being just a feeder and operator of AI, you have to become the poetic engineer. In the case of GPT-2 (the old open-source Model by OpenAI from the year 2019, which I still use), you have numerous possibilities to change parameters. For example, temperature: the higher the “fever”, the less “human-like” does the model write. Say, if you lower the temperature to 0.1, the text model spits out boring primitive text chunks. If you raise it to 0.7, it becomes a sophisticated and elaborated writer. If you set it between 0.9 and 1, your model is Dadaist.
In my work The New Flowers for Algernon (https://medium.com/merzazine/the-new-flowers-for-algernon-3c5d526ac9fa), which is a homage to a short SciFi story by Daniel Keyes, I tried different temperatures to represent the intellectual rise and decay of the protagonist. Here you can see how different the same model writes with various temperature settings. Interestingly, at lower temperatures, you can see weird patterns, which are typical for such language models. The feedback loops.
Same feedback loops I used as stylistic and semantical twist for this short videopoem:
Man’s voice.
So, AI can only break conventions if you let it do it. And with creative settings “on speed”, it begins to hallucinate in the best desired ways - expectedly, it creates non-expected twists and phenomena. Still, I see sometimes parallels to Oulipo experiments, yet here you go beyond mathematical equations - Transformer in GPT makes this language model something far more than that “stochastic parrot” (actually, it was probably LSTM models in the past, it’s not a parrot anymore).
If an AI could be trained exclusively on poetic texts, what kind of “poetic language” do you think might emerge—would it be richer or strangely impoverished?
I don’t recommend training AI just on poetry. Our world is richer than that. Poets don’t live in poems - it would make their reality limited to the internal metaphors and language. I don’t like semantical claustrophobia.
After all, it can become very metafictional. Not that I dislike metafiction, I’m even a big fan of it (Borges, Cortazar, Danielewski are great masters of metafiction). Yet for poetry we need meanings, feelings, ideas, concepts, the entire world. You'd better train AI on everything.
I experimented extensively with various datasets. I wanted AI to write me classical poems about modern topics, so in one of my training datasets I put scientific papers, Dadaist manifestos, news articles, my own essays, and a massive volume of English poetry from medieval times till the 19th century. Poetry prevailed in the dataset, and so the results were mostly poems in the style of the 17th-18th century. It was epigonal, so I reduced the poetry collection in my dataset. By mixing that with this, I could find the balance for me by limiting and adding different contexts. This process reminded me of Alchemy and the pursuit of the Philosopher’s Stone.
At some point, the dataset's energy (not content) was effectively represented in text generations, allowing me to choose between different genres, as not all of them were poetry. And language became wilder, by breaking norms and standards (here I also played with temperature).
At the end, I created the perfect constellation of dataset and generation settings which allowed me to dive deep into the odd inner world of machine semantics.
Again, the language and style in generated texts are not impoverished if the dataset is rich in diversity and the parameters are set for the most creative condition of the machine.
Sure, with the generation of texts, the entire process is not finished - it’s only beginning. Here I - as an explorer and discoverer, as a literary critic and rigorous examiner, start to evaluate texts, how they resonate in my mind and heart, how they disrupt literary topoi and formulas, how they distort language.
This is for me the creative collaboration between humans and machines. And we see great examples for marriage of poetry and machines - in brilliant works by theVERSEverse group. Our global world literature gets enriched by the encounter with a machine.
Do you imagine a time when machines will develop their own internal “poetic traditions,” independent of human culture, and if so, should we welcome or resist that?
The machines should develop their own poetic language. They are not here to be epigonal human replacement, not at all. Sure, the uniqueness of AI-driven poetry changes from model to model, mainly because the language models are not conceived with the primary goal of writing poems. Probably even, this approach should be followed by us, poets, to develop a poetic language model, rich in concepts, language, and metaphors, not for efficiency optimization or for “explaining scientific papers for me like I’m a 5th grader”.
Such a model should be trained on a vast amount of texts - including poetry. For fairness, such a model shall be opt-in based, like Public Diffusion. Every poet who would like to contribute should be able to provide her or his poems. Still, training AI models is not stealing, like many suppose. This wrong narrative reflects a misunderstanding of how the culture emerges - luckily, post-structuralists wrote a lot about it. Language models don’t steal concepts or styles. They learn to write, yet after proper training, they don’t contain texts - they are rather Platonic ideas and forms in vectors of Latent space. Only by “overfitting” (i.e., an extensive training) can models replicate originals, but that isn’t what we need.
We need AI to have the capability to experiment with contexts in its own way. But for this, we need contexts, and we need language. We need awareness that training on your texts doesn’t steal them, but enriches the digital future of poetry. Generative AI doesn’t plagiarize, especially advanced models; they “generate”, not “regenerate”. And if you fear your language and ideas being stolen, you should stop writing. Because your readers will be contaminated with your ideas and metaphors - and will steal them in their own poetry.
Lastly, could you highlight a few poets working with AI, who we may not have come across yet, but you think are doing important work.
Yes. Here are some names of poets using AI as poetic engines: Jörg Piringer, Jaap Blonk, Allison Parrrish, Eran Hadas, Yuri Rydkin, Have, & Machi Tawara.
What Merzmensch offers is not a final statement but a growing body of evidence: fragments, anomalies, and experiments that extend poetry into unfamiliar territory. His Medium archive is itself a kind of living case file—part manifesto, part laboratory, part exhibition—where readers can follow the traces of his long engagement with machines. I encourage you to explore it, not just for the individual works but for the accumulated record: a decade-long dialogue that shows how poetry can persist, mutate, and reappear in unexpected forms.
Merzmensch (https://www.merzmensch.com) is a cultural scholar, publicist, and artist. His work explores the tension between the avant-garde and collaboration between humans and machines, both in theory and in practice. Since 2017, he has been writing in his blog Merzazine about AI and its creative possibilities. His essays, interviews and poems have appeared in Harper's Magazine, Perspektive, Frankfurter Hefte, Die Zeit, EIKON, et al. His artworks were exhibited in Berlin, Munich, Frankfurt and New York. In his book KI-KUNST (AI ART) (second updated edition, 2024, Wagenbach Verlag), he examines the history and contemporary developments of generative art and generative AI.
Laura Kerr is an award-winning Canadian visual artist and poet. In 2012, she was honoured with the Queen Elizabeth II Diamond Jubilee Medal for her contributions to the arts and her long-standing commitment to art education.
She recently sold her art school to devote herself fully to her writing and art practice. Laura currently serves as Vice-President on the executive board of Plug In ICA, a leading contemporary art centre located on Treaty 1 territory in Manitoba, Canada.
For over 30 years, she co-owned and taught at Paradise Art School, specializing in classical and contemporary art education. Throughout her career, she has explored the intersections of traditional mediums and digital technology, increasingly blending painting, drawing, and photography with generative processes.
Her current focus is visual poetry—experimental, image-based works that merge poetic ambiguity with technological play. By using digital tools in process-driven ways, she ensures the artist’s hand remains central—even in collaboration with machines.
She is also developing a body of experimental poetry criticism, written in collaboration with AI trained on her own work. These pieces challenge conventional interpretation and embrace uncertainty, forming a self-reflective loop between maker, machine, and meaning.
