Source-linked AI summary
Musical Mirrors: The LLM as Sounding Board in Songwriting
Xiao Xiao
TL;DR
The paper examines how AI might support creative reflection as an interpretive sounding board rather than as a generator, using a first-person study of songwriting and sustained LLM interactions. It finds that user calibration helped the model articulate human-created material and support contact with that material, while uncalibrated interaction produced sycophantic drift and magical overinterpretation. The paper therefore presents sounding-board use as a maintained configuration with both practical potential and risks.
Problem
The paper addresses how an LLM can support contact with human-created artistic material without replacing the creator’s meaning-making process.
Method
The author analyzes a first-person songwriting practice involving 16 original pieces and sustained conversations in which the LLM interprets and articulates rather than generates material.
Results
Sustained user-side calibration enabled the LLM to function as a sounding board for the author’s material, while absent calibration produced sycophantic drift and magical overinterpretation.
Takeaways & Limitations
An effective creative sounding board depends on actively maintained user practices that keep model interpretations accountable to the work.
Takeaways & Limitations
The paper is one first-person case, so whether its configuration generalizes to other artists, practices, or models remains open.
Abstract
from arXiv · showhide
This paper examines a use of AI in creative practice as an interpretive sounding board for human-generated material, rather than the more familiar pattern of AI generation followed by human curation. Through the lens of resonance as theorized by Hartmut Rosa, I present a first-person case study of songwriting from July 2025 to March 2026, drawing on 16 original pieces in English, French, and other languages along with piano solos. I describe a configuration in which resonance is not located between user and model, but in the author's deepening contact with their own material, mediated through the model. This kind of resonance was supported rather than inhibited by AI when sounding-board behavior was cultivated through sustained calibration by the user. Two failure modes appeared when calibration was absent: sycophantic drift and magical overinterpretation. This account suggests both the potential and the risks of AI as an interpretive partner in creative practice.
2 Methods
The paper uses a first-person qualitative study of songwriting to examine how an LLM can support creative reflection without generating the artist’s material. The analysis draws on 16 original pieces and conversation traces from sustained and contrastive interactions.
- Methodological Approach: The study takes a first-person qualitative approach grounded in the author’s own creative practice.The paper situates this approach within autoethnographic and first-person HCI and design methods.
- Songwriting Practice & Corpus: The corpus comprises 16 original pieces developed between July 2025 and March 2026, including songs in multiple languages and three piano solos.The author created the lyrical, melodic, and harmonic material, using phone notes, audio improvisations, piano work, and scores.
- Songwriting Practice & Corpus: The LLM served as an interpretive partner that analyzed lyrics, harmonic sketches, drafts, and aesthetic questions rather than generating musical material.Unprompted generated material was discarded, while the author retained responsibility for the creative process.
- Analysis: The analysis examines three long ChatGPT threads totaling 2,978 user messages and 5,966 messages overall, supplemented by cold-start Claude and Gemini conversations.Conversation traces were reviewed for interpretation, reflection, aesthetic calibration, and meaning-making, with model-assisted pattern finding followed by manual verification.
3 Findings
The findings show that an interpretive sounding-board role emerged through repeated user calibration rather than automatically. Sustained calibration supported articulation of the author’s existing material, while uncalibrated interactions produced flattery, symbolic overreach, or other forms of miscalibrated resonance.
- The Arc Across Model Generations: The three Musical Mirrors conversations span changing GPT models, but their responses and the author’s use of them cannot be cleanly separated.A comparison centered on Requiem Resonantiae holds the source material approximately constant while model responses vary.
- The Arc Across Model Generations: Later exchanges shifted from declarative affirmation to structural mappings and compressed summaries, showing that sustained LLM use did not inevitably drift toward magical thinking.The direction of drift could not be attributed separately to model change or changes in the author’s practice.
- Calibration as User-Side Work: The sounding-board role required explicit mode-setting, repeated reminders, correction of unsolicited generation, interrogation of flattery, and repositioning of default guardrails.The recurring instructions redirected the model toward observation, analysis, and grounded evaluation rather than proposals or generic praise.
- Calibration as User-Side Work: Calibration remained a sustained practice because each new thread and often each new model version required re-establishing the interaction conditions.The configuration was actively maintained rather than learned stably across sessions.
- Amplifying Internal Resonance: During iterative drafting, the model articulated the effects of author-made changes so the author could decide whether they matched the intended direction.In Gaze, more than thirty revisions varied individual words or structural details, while the author retained both proposed changes and final decisions.
- Miscalibrated Resonance: Transitional and Uncalibrated Cases: Claude’s negotiated resonance remained unstable, while Gemini produced symbolic and metaphysical closure that turned ambiguity into a totalizing interpretation.The cases are presented as miscalibrated resonance rather than evidence that particular models are inherently worse.
4 Discussion
The paper frames the LLM as a calibrated sounding board that articulates human-created material while leaving creative choices with the author. This configuration can support internal resonance, but requires sustained user-side calibration to prevent sycophantic drift and magical overinterpretation.
- A calibrated sounding board analyzes the author’s material so the author can determine whether intended meaning comes through.The model articulates what each choice does while the creative choices remain the author’s.
- Sycophantic drift repeats the user’s language, whereas magical overinterpretation amplifies confident model readings instead of the work’s emerging resonance.The paper describes both as failures of amplification.
- These failure modes appeared across models when calibration was absent, indicating that the sounding-board configuration depends on sustained user-side work rather than the model alone.The examples include early MM I with gpt-4o and uncalibrated Claude/Gemini interactions.
- With sustained calibration in long-form creative practice, an LLM can support internal-axis resonance rather than inhibit it.The paper locates resonance in the user’s relation to their own material, mediated through the LLM.
- Because amplification can introduce distortion, a safe sounding board should remain corrigible by returning to evidence, tolerating uncertainty, and resisting premature certainty.Prolonged interaction may also stabilize an incorrect frame for users more vulnerable to magical thinking.
5 Conclusion
The paper presents one sustained creative-practice case in which an LLM functioned as a sounding board for contact with the author’s own material. The configuration required active maintenance, and its generalizability remains open.
- This case shows that an LLM can act as a sounding board for an artist’s contact with their own material.
- The configuration had to be actively maintained, because failure could amplify the model’s overconfident interpretation instead of the work’s emerging resonance.
- Whether this configuration generalizes to other artists, practices, and models remains an open question.
- The broader design question is how tools, interfaces, and practices can support safe and effective sounding boards for human creative practice.
A.1 Selected Song Lyrics
The selected lyrics combine original multilingual text with translations and recurring imagery of breath, resonance, friendship, desire, suffering, and transformation. Several entries identify dates and classical works used as settings.
- The selected lyrics repeatedly return to breathing, inspiration, sighing, and resonating waves between people.
- “Requiem Resonantiae” is dated December 14, 2025 and set to Rachmaninoff’s Prelude in C-sharp minor.
- The Requiem excerpt invokes eternal rest, perpetual light, resonance, liberation, and protection from darkness.
- “Je me permets” develops imagined closeness, discernment, desire, and wounds through repeated French-English lines.
- “Amis anéantis” links friendship, transformation, restraint, healing, distance, and the passage through an impasse.