Shared agency means that no single participant—not a dancer, not the group, and not the AI system—fully determines what happens next. In an AI-supported dance cypher, the useful question is therefore not whether the machine “made the music,” but whether its responses help people notice, influence, and coordinate with one another.

Agency is more than pressing a button

Agency is the felt and practical capacity to affect an unfolding situation. A conventional music player gives a dancer control over obvious actions such as play, pause, and track selection. A generative system can create a looser relationship: movement changes the music, the music changes the movement, and each participant adjusts to the others.

That loop becomes shared agency when the result cannot be traced to one isolated command. The term does not mean that the AI has human intention or responsibility. It describes how control is distributed across people, sensing, interface rules, and generated output.

A dance cypher makes the social loop visible

A cypher is a circle in which dancers take turns entering the centre while the group watches, responds, and sustains the event. It is not simply a sequence of solos. Attention passes between the person dancing, the people at the edge, the music, and the conventions of the gathering.

On 16 September 2026, Zhixing Chen and Cheng-Zhi Anna Huang released the arXiv preprint Encypher: Shared Agency and Social Presence in Collaborative Music Generation for Dance Cyphers. Encypher translates qualities of collective movement into text prompts that condition real-time music generation. Instead of giving one person a controller, it uses the room’s movement as part of the musical input.

The project was developed through five weeks of co-design with local dancers, then examined through a study with previously unacquainted participants, a public museum event, and a live performance. The authors report that participants described the music as responding to the room’s energy and developed a sense of shared agency. Newcomers sometimes felt uncertain, but that uncertainty also prompted them to look to other dancers for cues.

The system does not need to understand the dance as a person would

It is tempting to say that the AI “feels the room.” That metaphor goes too far. A movement-analysis pipeline detects selected qualities, converts them into prompts, and generates music within design constraints. It does not have the situated history, vulnerability, or cultural knowledge of the dancers.

Yet the system can still reorganise attention. Imagine the music becoming denser when several people move with greater collective intensity. A dancer may hear the change, notice who else is moving, and answer with a pause. The socially important event is not hidden inside the model. It is the new chain of attention among people around it.

This differs from a solo body-to-sound instrument. Somatic-AI Lab’s earlier discussion of sound-conditioned movement generation examined how one modality can guide another. Encypher asks a more social question: can a generative mapping help a group coordinate without appointing a single operator?

How can we tell whether agency is genuinely shared?

Three observations are more useful than asking whether the music sounds impressive.

First, watch where attention travels. Do participants look only toward a screen, or do they look toward one another? Second, test whether different people can alter the unfolding result, including quieter participants at the edge. Third, examine breakdowns. If newcomers cannot tell what the system is responding to, uncertainty may support social cue-seeking—or it may simply exclude them.

These questions make evaluation partly technical and partly social. Latency and mapping consistency matter because delayed or arbitrary feedback weakens the loop. But interviews, observation, and co-design are needed to learn whether participants actually experience influence, presence, or exclusion.

Shared agency does not remove design responsibility

Calling an experience collaborative can conceal who chose the sensors, movement categories, musical model, and acceptable forms of participation. Designers still determine which bodies are legible to the system and which actions produce a response. Shared agency is therefore an achievement to investigate, not a label to apply automatically.

The practical value of Encypher is its shift in emphasis: generative AI for dance can be evaluated by the encounters it helps organise, not only by the media it produces. This week’s community report on uncertain and actively gathered evidence offers a parallel lesson from robotics: a system should make clear what it sensed, what it inferred, and how those choices shaped the next action.

References

  • Chen, Z., & Huang, C.-Z. A. (2026). Encypher: Shared agency and social presence in collaborative music generation for dance cyphers. arXiv. https://arxiv.org/abs/2609.18062