Abstract Human–AI interaction is typically organized as a conversational loop: humans formulate natural-language prompts, which are tokenized and encoded as numerical inputs to a machine learning model. While input and output appear as symbolic exchange, the model’s internal processing is subsymbolic, relying on embeddings, attention mechanisms, and distributed representations across a neural network. The system produces a response that humans assess and iteratively refine, enacting a paradigm in which meaning-making is organized around human semantic control and evaluation.
This contribution intervenes in that loop by foregrounding the body as an epistemic site and shifting interaction from prompting to parametric modulation. Drawing on the Japanese dance practice of Butoh, it proposes a lecture-performance in which bodily states—breath, stillness, tension, balance—are sensed and used to modulate generative parameters of a large language model. Rather than altering semantic content, these signals influence conditions of generation (e.g., temperature, latency, output length, memory decay), reshaping the model’s distribution of possible outputs while leaving the prompt unchanged.
Informed by new materialism, the experiment reframes human–AI interaction as embodied co-becoming rather than transparent tool use. The body is neither reduced to an input device nor positioned as commanding the system; instead, interaction unfolds through delay, uncertainty, and partial responsiveness. Attending to pre-reflective sensation and indeterminacy, the work values fragmented or unstable outputs not as errors but as traces of subsymbolic dynamics. It thus demonstrates how parametric modulation can engage generative AI beyond language, opening space for affective, non-verbal forms of knowing within and alongside computational systems.