Adversarial Poetry
Critical Humanities and the Limits of Artificial Intelligence
Symposium, 1-2 October 2026, Venice

Poetry in the history of language formalism and information theory

Poetic experiments with the formal properties of language suggest that LLMs combinatorial techniques belong to a much older history. Medieval Arabic prosody and cryptography developed formal procedures for generating and classifying linguistic structures, anticipating later experiments in algorithmic and constrained writing. In the 19th century, Ada Lovelace saw in the Analytical Engine, a prototype of the digital computer, the offspring of a “poetical science” (Toole 1992). While for the Russian Formalists, poetic verse was a privileged field for the application of statistical methods, insofar as it could be analysed through patterns of rhythm and breath in human language. This privilege was overturned in the late 1940s, when Shannon’s information theory regarded poetry as one of the most difficult messages to encode in a communication channel, owing to its high degree of unpredictability. Starting from the 1960s, numerous experiments exploited the combinatorial potential of the computer, developing algorithmic and procedural writing, and finding in the alienation of language through the machine the condition for a new poetic ethos. Since then, poets were no longer the vates of the nation, in its Romantic tradition, nor of its language.

Poetry and rhetorical tropes are ways to conceal or displace meaning. It would be limiting, however, to ascribe these functions exclusively to what cultural critique designates as “artistic texts.” Obfuscation and scrambling are at work in every secret language or “dark tongue” (Heller-Roazen 2013). While the automation of language is but a chapter in the longer history of utopian artificial languages and the dream of a single perfect idiom (Eco 1997), human communication is in fact founded upon misconceptions, hegemonic and subaltern uses of language. The friction between the aspiration to linguistic immediacy and the irreducible opacity of encoded and embedded meanings has acquired new significance with the emergence of LLMs. These systems offer a novel terrain for testing semiotic and narratological hypotheses, yet they also transform every linguistic phenomenon into a commodified production and exchange of tokens.

Icaro Lab: Poetry as a logical limit of AI and technique of jailbreaking

Recent research conducted by Icaro Lab has demonstrated that poetic prompts can jailbreak LLMs, suggesting an adversarial function of poetry within computational systems (Bisconti et al. 2025). The same research group has further investigated LLMs safety by drawing on narrative roles derived from Propp’s Morphology of the Folktale (Bisconti et al. 2026). This line of inquiry has given rise to a new field of research and methodology termed adversarial humanities (Galisai et al. 2026). This framework is designed to test LLMs safety against diverse forms of rhetorical manipulation in prompting, with a particular focus on stylistic obfuscation and goal concealment. Collectively, this research indicates that rhetorical and poetic forms of manipulation function as a privileged tool for probing and destabilizing model safety constraints.

That poetry may represent a limiting case for AI is also suggested by current debates concerning LLMs capacity to produce and comprehend irony and ambiguity (Mazzoli et al. 2025; Leivada et al. 2023; Lewis & Mitchell 2024). While such models have been conceptualized as ideological or poetic machines (Weatherby 2025), they operate within an autoreferential structure in which words refer primarily to other words and generate recognizable “genre” outputs. It remains unclear whether they merely extract and recombine metaphors and analogies from the repository of human text, or whether they can genuinely instantiate novel figurative relations. This uncertainty encourages methodological approaches that seek to employ various entropy measures (informational, phonological, semantic, etc.) to evaluate the utility of different databases such as poetic texts, narrative prose, code, legal discourse, encyclopedic writing, etc. (Gambetta et al. 2025a; 2025b). Poetic and narrative texts, for instance, could help mitigate model collapse: their high semantic entropy may reduce the degradation and loss of diversity typical of flatter, more conventional textual forms. This proves, once again, that AI requires datasets of high-quality knowledge (namely well-written books and academic publications) that cannot be found simply scraping internet sources.

Critical humanities in AI studies, AI literacy, and STEM disciplines

These technical phenomena are in need of explanation: we argue that the humanities are in a position not only to trace the history of such problems in linguistics and semiotics, but also to provide a conceptual and operative toolbox to address contemporary challenges. The idea behind the symposium is indeed to stage an encounter (probably a clash) between engineering and literary form, between critical humanities and STEM disciplines.

Disciplines such as digital humanities have long engaged with the impact of computational technologies on the understanding of culture, for example through methods such as distant reading. Building on this trajectory, the symposium aims to address the influence of poetic and rhetorical structures on deep learning systems. It not only considers the challenges posed to humanities curricula by the introduction of AI, but also seeks to show how humanities can intervene in the widespread and often uncritical deployment of AI also in education, highlighting unresolved questions that cannot be reduced, as is often the case in current approaches, to the design of  ethical frameworks alone.

Ultimately, rather than speaking of the simple digitalization of culture, these developments should be understood in terms of a dialectical relation in which cultural forms and computational systems mutually shape and transform one another. No AI curriculum, therefore, should be organized without sustained engagement with the history of logic, semiotics, rhetoric, linguistics, and digital humanities.