Movement out the other end
21 November 2026
This past Monday, after a long weekend of researching and digging deep into my work, I sat down with my team to discuss the architecture of building a movement model. These meetings are usually intense, because we passionately want to do the right thing, whatever that is, and because of the excitement of collaborating across disciplines. That last part is important. A dancer's movement mind is very different from an engineer's. I am the dancer, and fortunately or unfortunately, my mind jumps. It moves by association, from the body to a sound and back again, often before I have finished a sentence. The engineers catch those leaps and give them a form precise enough to be built and tested. So I truly appreciate this team for putting up with my multifaceted way of thinking, and for turning it into something we can run. But something struck me during the meeting, and I had to write it all down, as it was more of a revelation than what the team needed to discuss.
Driving home, I couldn't stop thinking about why AI generating movement is seen as something terrible. Perhaps I am biased, as I am the one building it with this team, and that perspective matters. In any research, you look at what others have built and follow their footsteps so long as you believe they are right and their work shares much of what you are building. One of the engineers had researched the scholarship and found benchmarks from groups that have built GPT-style models for dance specifically, and they were all interesting. I read some of this work years ago, and at the time it was as exciting as I find it now, and as my students find it today. What stopped me in my tracks was the idea of replication. Many AI systems for dance are geared towards the product, just as much of the world works. Can we produce a dance? Will it entertain as well as the best human dancers would? Is it believable? Is it a real dance? Can we replicate human movement so closely that people cannot tell an AI-generated dance from a human performance?
One of my students asked me directly, "What do you want to come out of the other end of this architecture and pipeline?" I responded, "I want movement to come out the other end." Confused, the student asked, "So, not dance?" I said, "Yes, not dance." I want a body that can move, and a system that allows the person using it to create with it. That answer already breaks the chain of AI production, because with AI you only need to send it on a task, and it will produce the materials, including whatever else it decides to bring with it.
The questions about replication may not have entered the room during the build phase, but the assumptions were already in the same room. To make a good model, we need to resemble humans as best we can. That perception has driven many builders of AI for movement into what has now become a stumbling block between AI innovation and human practice. People are right to condemn AI-produced dances, because we know AI cannot dance.
As a dance practitioner myself, my work follows the school of thought that says a dance is made, not purchased. For a dance to be considered a dance, it has to be made from a conscious place. A decision is made about what is being created and a body is trained to accomplish it. Then come the iterations, which are often where the artistry is born. Conscious does not have to mean planned. A dancer improvising, or answering a rhythm they have known since childhood, is choosing through their training and their memory, even when no one would call it deliberation. This happens at every level of dancing. A dancer in a nightclub has to prepare before arriving at the dance, and then makes choices about what to dance, or at least how to follow the song. These choices cannot be left out of making a dance, or out of dancing itself. When you take that piece out, you get what we now see as AI-slop choreography. No thinking or artistic iteration goes into it, and no relationship feeds back into the work being dished out. Even worse, the thing that is dancing is a mimicry of something we know to be familiar and better.
After this thought, it dawned on me why everyone around me has been confused that I would be the one building this kind of venture, because this has been the shady cloud over my research into building a movement model, an AI system for human movement. It is not yet time to discuss what is actively being built, but I think we could use some explanation of what the vision is, and what it is not.
Popular AI systems are trained on large sets of data, and a diverse dataset is usually more desirable. So builders of AI systems for movement collect large amounts of data in order to achieve a fully functioning movement AI. Our model learns from recorded movement too, and it is meant to. What I am building in the Movement Language Model is an AI system for dance that supports artistic creation and ideation and gives dancers more tools to support their work.
So, what is this movement I refer to? I made the case that a body that comes out of the model should understand weight, shape, effort, pedestrian movement, dynamics, dimensions and levels, and also have a sense of sound, timing, breath and rhythm at its scales. This means that training this model is more than assembling large amounts of movement data. Alongside the movement, we are training it with many other kinds of information, such as sounds from around the world, geographical coordinates, light, gravitational pull, pressure and muscle articulation. We read all of this through units we call emes, including the kineme, which is visible movement, the someme, the felt interior of effort, the soneme, which is sound, and the tereme, the land. The kineme and the tereme are trained and the soneme is under test. The someme is still being built. The model is beholden to the person using it. It is not trained to reproduce a particular person's movement, or to take ideas from people whose movements are already published to the world. It offers movement that an artist can question, take apart, choose or refuse, and develop, and it works from the artist's own material. The decisions that turn that movement into a dance stay with the artist, and so does the freedom to make the work their own. The dancers whose recorded movement the model learns from are asked first and credited for what they give.
This is the task at hand, and it is enormous. I could disturb you with the architecture of the model, but I will not, for many reasons. One of them is that we are still experimenting with ideas, and our goal is to do something new. We are building the first culturally conditioned AI system. I left Monday's meeting grateful that the room where this is being built holds both the dancer and the engineers. I believe that we can do this ethically and properly, and I hope that the movement world will join us and develop this model that is made for humans and by the communities that continue to sustain this embodied knowledge that is primordially given to us all. We plan to open a beta in 2027, and I would love for you to be part of it.