Selective legibility: From material readout to machine agency
Materiomusic starts with a reversal: listening becomes a way to know matter. In Markus J. Buehler’s account, molecular vibrations, spider-web structures, flame dynamics, and material failure can be translated into musical form through relation-preserving mappings. The point is analytical: hidden structure becomes perceptible and composition can feed back into material design.1 This commentary asks whether embodied AI needs a civic counterpart: reading machine agency through constrained public readout.
The argument exceeds sonification. Mapping temperature to pitch or genomes to melody can help, but materiomusic asks for stricter source-constrained translation. A protein’s vibrations are constrained by its chemical bonds, a web carries tension through geometry, and a flame responds to acoustic forcing. Music becomes a constrained readout when another medium gives access to hidden structure without erasing key relations. Buehler calls this an epistemic inversion: listening becomes a mode of seeing.1
Buehler develops selective imperfection as a generation under constraint.1 Rich behavior often sits between sterile order and unreadable disorder. Here, the analogy is normative about readout; it does not claim that proteins, webs, flames, and AI systems share the same structure or social stakes. Since selectivity is unavoidable, design must make it accountable. Public readouts should simplify honestly, preserve judgment-relevant relations, and acknowledge loss. This aligns with the demand that AI-guided scientific claims remain transparent, explainable, challengeable, and subject to human judgment.2 Public AI needs an account of legibility beyond the opacity-transparency continuum.
Selective legibility offers one such account: grounded, partial, contestable readout conditions that make machine states, intentions, guidance cues, traces, and consequences publicly interpretable. Readout regimes define when one version of a system’s state becomes the version people use.
Ordinary transparency sets a different expectation. Even open systems may offer logs, scores, explanations, or traces unreadable to most people. Selective legibility asks what should be readable, for whom, through what medium, at what moment, and with what admission of loss.
The black-box metaphor is too narrow for public AI because it concentrates attention inside the machine. Meaning in embodied systems emerges across policy, motion, interface, observer, replay, and memory. People infer intention before an action is complete; they may read accurately, overread, or be guided by design.
Consider an autonomous shuttle approaching a crosswalk. It senses a pedestrian, predicts motion, slows or stops, activates a light or display, and leaves sensor traces and logs. The pedestrian needs orientation, operators need a diagnosis, and regulators may need a later record. The encounter can be read through five layers: state, intent, guidance, evidence, and write back.
