Structural and Representational Agency in LLM-Driven Timbre Generation and Interpolation
The study investigates how agency—the ability to cause an effect versus knowing that you caused it—is separated in AI-driven sound generation. It shows that structural constraints survive model failures, meaning the physical limits of what can be generated remain consistent even if the underlying model is flawed. However, guarantees based on the model's own reports are fragile and fail when the model malfunctions.