Recursive Misinterpretation II

Recursive Misinterpretation II Carlos G. Román Experimental Music & Digital Media Louisiana State University cr[email protected] Roberto Mochetti College of Music & Dramatic Arts Louisiana State University [email protected] Jesse Allison Experimental Music & Digital Media Louisiana State University [email protected] Abstract This participatory performance piece builds on a series of process-based compositions exploring the creative tension between human intention and AI-generated outcomes in text-to-audio generation. The original work employs a recursive structure where prompts guide alternating contributions from a human composer and a generative AI model, embracing misinterpretations as creative drivers. For this iteration, the process is extended into a live, distributive format incorporating real-time audience participation. Audience members submit textual descriptions, which are synthesized by a large language model into a single prompt that is then used to generate new audio via a text-to-sound model. This AI output becomes the foundation for the next cycle of prompts, creating a feedback loop between human perception, language interpretation, and machine generation. The piece emphasizes the artistic value of semantic gaps and misalignments in AI interaction, proposing generative miscommunication between humans and AI models as fertile ground for collaborative sonic exploration. 1 Origins and conception of the piece This distributive performance piece originates from an idea explored in a series of works called Recursive Misinterpretation (2025) by Brazilian composer Roberto Mochetti, which were initially structured as process pieces, guided by a set of instructions rather than a fixed score. The conceptual basis of these pieces explores the interpretive gap between human intention and machine-generated outcomes in text-to-audio generation. In the piece, the composer and an AI model alternate in generating musical movements. Starting from a simple technical prompt, the composer creates a first section, then uses the prompt to generate audio via an AI model. The composer then listens to the AI output and writes a new prompt based on their interpretation of it—this guides the next human-composed movement. The process repeats, with each new prompt derived from the previous AI output, continuing until the results feel complete or the AI output becomes uninteresting. The piece ends with an AI-generated section. Therefore, the initial prompt serves not as a blueprint, but as a point of departure for iterative, dialogic creation. This iterative use of prompts has been employed as a process of exploration and experimentation in different areas of knowledge (e.g. Hutson & Cotroneo, 2023), allowing non-expert users to better grasp the creative potential and boundaries of generative AI models. For this conference, we propose extending this composition process into a live, interactive format where the audience contributes prompts in real-time through a web interface. We are also inspired by the integration of multimodal generative tools in different scenarios (e.g. Sturm et al, 2024). Firstly, audience members describe the initial fragment of a potential piece in their own words. These prompts are not fed directly into the generative model, instead, all submitted prompts are Proceedings of the 6th Conference on AI Music Creativity (AIMC 2025), Brussels, Belgium, September 10th-12th sent to an open large language model (LLM), which summarizes and synthesizes them into a single coherent prompt that describes what the consensus of the audience perceives at a particular moment. This collective prompt is then sent to the text-to-sound open model to generate a new audio piece, which gets mixed in real-time with the previous fragment. The prompts submitted and the collective prompts generated are also visualized and projected in real-time to the audience. The process can be repeated iteratively: the newly generated piece becomes the basis for another round of audience prompts, forming a feedback human-in-the-loop between human description, LLM interpretation, and AI sound generation. This twist opens possibilities for unpredictability, authorship dispersion, and collective creativity. For the conference, we propose an 8-to-10-minute version of the piece. 2 Thoughts about the process and role of the composer/audience The intention with this distributive performance is that the audience interacts with AI generative tools (text-to-music models), using structured prompts to elicit musical material. Rather than correcting or rejecting AI “errors” or “miscommunications” in the generation, the process embraces them: divergences and unexpected sonic results are interpreted as creative opportunities to move the piece forward. These “misdirections” between the prompt and the sonic outcome are recursively integrated into the composition, informing the next instruction or section of the piece. The composer and the audience assume a hybrid role as authors, editors and curators of the AI generated pieces, engaging in a feedback loop with the AI system. Another key question explored through this set of process pieces: Can we navigate and meaningfully understand the latent space of a text-to-music model using prompts alone? Central to the proposed performance is the acknowledgment of a semantic gap (Allison et al, 2024) between the audience's intention, the written prompt, and the AI's audio output. We also want to highlight the contrast between utilitarian uses of generative AI music tools and experimental, process-based creative practices like the one presented in this piece. This piece builds on top of the original work by using generative AI and prompts to allow a synchronous collaboration between audience participants, instead of the AI/composer collaboration explored in the initial piece. 3 Conclusion This performance highlights the generative potential of human-AI misalignment in creative processes. It aims at demonstrating how AI text-to-music models can function less as a tool for realizing fixed ideas and more as a collaborator for sonic exploration in an evolving artistic dialogue. Because of the limitations of current AI generative models, different creative approaches can be used to add a new dimension to the human-in-the-loop dialogue. The methodology suggests possibilities for collaborative creation or distributed authorship in future performances. The approach invites reflection on authorship, control, and interpretation in the age of generative music technologies. Acknowledgments References Allison, J., Farrar, D., Nash, T., Román, C. G., Weeks, M., & Ju, F. X. (2024). Play Me Something Icy: Practical Challenges, Explainability and the Semantic Gap in Generative AI Music. Proceedings of the 2nd International Workshop on eXplainable AI for the Arts, ACM Creativity and Cognition Conference, Chicago, Illinois, June 23, 2024 arXiv preprint arXiv:2408.07224. https://arxiv.org/pdf/2408.07224 Hutson, J., & Cotroneo, P. (2023). Generative AI tools in art education: Exploring prompt engineering and iterative processes for enhanced creativity. Metaverse, 4(1). Sturm, B., Amerotti, M., Dalmazzo, D., Cros Vila, L., Casini, L., & Kanhov, E. (2024). Stochastic Pirate Radio (KSPR): Generative AI applied to simulate commercial radio. In Proceedings of the 5th Conference on AI Music Creativity, Oxford, England, 2024. 2