Conflicts around the role of artificial intelligence in culture are already raging in full swing, but they don’t directly affect poetry yet. Nevertheless, the first experiments are already underway: it has been reliably established that in blind experiments, general readers prefer machine-generated imitations to real Byron and Whitman, perceiving them as more authentic. This is rather good news than bad: imitability is one of the signs of a truly powerful authorial individuality. However, it’s clearly time to ask ourselves questions about this.
Questions
1. If, in the end, well-developed AI enriches us with new poems by Rimbaud and Brodsky, nothing will be lost from the dead classics, right? Or is extra-textual authenticity more important than the text’s correspondence to the most qualified expectations of the reader and even the researcher? And isn’t this possibility fraught with difficult-to-resolve epistemological and ethical problems?
2. There’s no objection to working authors using AI as a tool that expands the horizons of their own poetics — because any tool in an author’s hands becomes part of the teleology of their creative world. But surely no “writing machine” in the figurative sense can keep up with a literal writing machine? What are the competitive chances of an author working the old-fashioned way alongside AI’s multiplying potential?
3. The turning point in poetry-AI relations will come when new technology creates a new convincing authorial individuality (obviously not identical to the authorial individuality of the human behind this project, if they had one). What cards might remain in the hands of those wishing to compose from themselves in this case?
Responses
1. It seems to me that this question proceeds from an implicit assumption that there will be many (some noticeable number of) people wanting to compose poems “in the voice of” existing authors, dead or living, with the help of large language models (LLMs). But this isn’t necessarily the case: naturally, in the first stage, when boundaries of the possible are being demarcated, we see many experiments precisely in this area: can a model write a text that will be accepted by an expert as an unpublished poem by Alexander Pope? Can a model write another hundred and fifty Shakespearean sonnets of such quality that, hypothetically speaking, a non-professional reader couldn’t distinguish the “real” from the “fake”?
Just recently, however, we witnessed how in the process of explosive development of TTI (Text-to-image) models, like Midjourney, a technique of peculiar stylistic pastiche gained and quite quickly lost enormous popularity: “Burning Man festival in 1920s photographs,” “family selfie in the style of Studio Ghibli/Flemish school/El Lissitzky.” After we begin to more or less understand the tool’s capabilities, we inevitably think about how else it can be used: aren’t there more interesting techniques than imitating styles and mixing them?
In principle, the question of what’s more important — “extra-textual authenticity” or correspondence to expectations — reduces to the question of why we read. If we read to get certain impressions, to evoke certain emotions related to the reading process, then apparently correspondence to expectations is more important. If the goal of reading is to gain new knowledge about unique human experience, then “extra-textual authenticity” is more important.
At the same time, the question of the epistemological status of certain artifacts generated by generative AI is, of course, theoretically very interesting. Since AI (so far) has no biography, history, and specific position in the sociocultural system, and also lacks agency and intentionality (wants to tell us nothing), a gap emerges between the text’s surface and its ontology — i.e., where it came from and what it “has the right” to communicate. And if it says nothing about a person, time, history, or unique inner experience — then what does it say? Can we consider that a text devoid of contextual/historical foundations contains any knowledge at all, and if so, what kind? We find ourselves face to face with a text about which we know nothing. Its origin is unknown to us, but since (for now) LLMs function as black boxes — we have no detailed understanding of how they do what they do — we can only guess what influences and sources are imprinted in the text, where the stylistic register comes from, and so on. How (and why) should we interpret a simulacra text written by a model “in the voice of” Alexander Pope? If we gain some knowledge from this interpretation, then what is it about — and to what degree? About Alexander Pope? About a “model of Alexander Pope”? About the large language model? About the English language of the first half of the 18th century as represented in written sources included in the training dataset?
2–3. The answer to these questions, it seems to me, makes sense to combine. I wouldn’t want to indulge in speculation about potential future capabilities, so I’ll speak about AI — more precisely about LLMs, since those are what interest us here — in the form they currently exist, or in such an improved form that can be linearly extrapolated from the present. So, in this form, as I see it, models are most productively viewed within the theoretical framework of extended cognition. In this sense, the only somewhat close — and yet very imprecise — historical analogy to what’s happening turns out to be the emergence of writing. Both radically expand cognitive boundaries, working, of course, in completely different ways.
We can imagine how poetry changed with the emergence of writing: expansion of themes, liberation from mnemonic functionality (or its strong weakening), formal complication, transition from anonymity to authorship, finally, the possibility of effectively accumulating tradition, as a result of which the trajectory of future development became more determined by the totality of past states. What happened in the properly cognitive sense? The possibility emerged for stabilization and external fixation of thought. Writing allows mental contents to be carried beyond the brain’s boundaries — fixing, transmitting, and accumulating knowledge across time and space, thereby expanding memory and rational control. Two important properties of writing should be noted. First is its weak interactivity: in first approximation, writing only preserves but doesn’t respond or process information. This means writing contributes to long-term cognitive architecture without changing much in “real-time” processes. Mainly, writing is directed outward and toward the future; it creates the possibility of thinking in delayed perspective and ensures cultural continuity, but doesn’t help think “here and now.”
The cognitive extension that LLMs bring has a different nature and direction. LLMs don’t simply preserve or transmit thoughts, but directly co-participate in their generation (offering variations, associations, arguments, etc.), thereby expanding our generative and linguistic abilities and acting as a simulative interlocutor with an alternative line of thinking. LLMs, unlike writing, have high interactivity, are not a static cognitive extension (like writing) but a dynamic one. LLMs perform the role of a peculiar technological prosthesis of intuition and expand linguistic imagination. Unlike writing, directed outward and toward the future, LLMs as cognitive extension are directed inward, deeper — and toward the present. The model embeds itself in the thinking process as, well, let’s say, a second/alternative “inner voice” capable of suggesting unexpected cognitive moves.
Now, keeping in mind the changes that occurred in poetry with the emergence of writing, one can try to imagine how it will change at this new stage of cognitive technological evolution. I must honestly admit that I lack the imagination for this — but it seems that many things that have value for us in poetry and its practice will apparently be devalued — but new ones will emerge, just as happened during the transition connected with the emergence of writing.
I don’t think that in this new world the question of “productivity,” as posed in the survey, will make sense — and the opposition between “composing from oneself” and composing using LLMs will most likely not hold up. In its current form — definitely not.
Author on LinkedIn.