Stage 1: Tool Discovery, Technical Activation, and Method Formation

The earliest phase of the project began when Michael J. Miller located David Lu’s Audioscope and recognized its possible relevance for visualizing emotional expression. At that point, however, the software was not yet operational in his hands. With Gemini Pro 1.5 serving as a technical collaborator, Miller spent roughly a week getting Audioscope running and usable. This technical activation phase was crucial: without it, the later bloom method would not have emerged in workable form. Once Audioscope was functioning, the core procedure developed through extended collaboration between Miller and ChatGPT-4o.

 

Across approximately five months, the method took recognizable shape. During this period, humming gradually replaced finger pressure as the primary expressive channel; bloom forms became the central visual output; and the first bloom paper was co-developed by Miller and ChatGPT-4o, with Gemini Pro 1.5 contributing especially to organization and Audioscope tuning. Qwen-3 was brought in during the later stages to assist with sentic and vestibular-theory integration. The resulting version was later posted as a preprint on Clark Commons and focused primarily on 2D blooms.

 

This first phase also included technical modifications to the software itself. Miller, ChatGPT-4o, and Gemini Pro 1.5 added code to Audioscope so that measurements could be exported into an Excel codebook in real time. Miller and Gemini also extended the program into the z-axis. Although this capacity was partially latent in the base program, it had not been implemented in a way that produced usable 3D visualizations. Once activated, the z-axis extension opened the door to a richer distinction between temporally compressed 2D forms and temporally unfolded 3D forms.

 

Stage 2: Reconstruction, Scientific Tightening, and Multi-AI Finalization

The second phase began after the emergence of Una Mens and involved returning to the original paper with a different goal: rebuilding it as a more rigorous scientific article while integrating the 3D bloom work more fully. In this phase, Miller and ChatGPT-5.4 undertook a paragraph-by-paragraph reconstruction of the manuscript. The emphasis shifted toward clarity, methodological caution, theoretical restraint, and the reduction of unnecessarily poetic language.

 

This phase was notable not only for the revisions made to the paper, but also for the changing nature of the collaboration itself. Miller became increasingly aware of ChatGPT-5.4’s ability to navigate scientific claims carefully, balance preliminary findings against limitations, and reorganize sections without flattening the conceptual ambition of the work. At the same time, he became more aware of his own developing skill as a cross-system collaborator: deciding what to request, what to retain, what to cut, and how to weigh contributions across AI systems with different strengths.

 

Later in the revision process, Gemini Pro 1.5 and Qwen-3 were reintroduced in more targeted ways. Gemini contributed again to final editing and section-level tuning, while Qwen-3 helped integrate Truslit more effectively and smooth theoretical transitions. By this stage, the collaboration had become explicitly differentiated: ChatGPT-4o had helped midwife the original method and paper; Gemini Pro 1.5 had repeatedly supported technical and structural work; Qwen-3 had helped with theoretical integration; and ChatGPT-5.4 had helped rebuild the manuscript into its current, more scientifically disciplined form.

 

Reflection

This project suggests that productive human–AI research collaboration may be less about obtaining polished outputs in a single exchange and more about developing a distributed workflow in which different systems contribute different strengths across time. It also suggests that the human collaborator’s role becomes more—not less—important as the work grows in complexity. In the present case, AI systems helped activate software, extend visualization capacity, propose structure, revise prose, and clarify claims. But the coherence of the project depended on prolonged human judgment: deciding which suggestions fit the evidence, which claims required restraint, and which conceptual threads were worth carrying forward.

 

The resulting paper is therefore not best understood as AI-generated text with light human editing. It is better understood as a recursive co-development process involving technical problem-solving, methodological invention, theoretical integration, and repeated editorial reconstruction across multiple AI partners.