Somatic-AI Content Platform

Your Body Knows Things the Algorithm Doesn't — Yet

A Practitioner's Guide to Generative AI: What It Can Do, What It Can't, and What's Worth Your Attention

You have probably seen the videos. A skeleton glides across a screen, conjured from a text prompt. A dancer's silhouette is captured by a phone camera and instantly remapped into a shimmering avatar. Motion-capture data that would have required a Hollywood studio five years ago is now generated in seconds on a laptop.

The hype is real. So are the limitations. And if you practice yoga, Feldenkrais, somatic experiencing, Body-Mind Centering, contact improvisation, or any movement discipline that takes interoception seriously, there is something specific you need to understand before you either dismiss this technology entirely or surrender your discernment to it.

Here is a clear-eyed map.

What Generative AI Can Actually Do Right Now

It can synthesise plausible movement. Text-to-motion models — trained on enormous libraries of motion-capture data — can now generate skeletal animations from written descriptions. "A person rises slowly from the floor and turns toward a window" produces something that looks, at a glance, like movement. These models have improved dramatically; the uncanny stiffness of early results is giving way to something more fluid. The key word, though, is plausible. The system has learned statistical patterns of what movement tends to look like. It has no body.

It can track and respond to your pose in near-real-time. Pose estimation tools — the technology running quietly inside everything from fitness apps to interactive art installations — can identify thirty-three body landmarks from a standard phone camera, at frame rates fast enough to feel responsive. Systems built on this foundation can trigger sounds, shift visuals, or generate reactive environments based on your position in space. This is genuinely interesting for movement artists and researchers.

It can build visual worlds from a skeleton. Image-generation models, conditioned on pose data, can clothe a detected skeleton in an infinite variety of visual styles: ink wash painting, bioluminescent field, architectural drawing. Your movement becomes a generative brush. Several artist-researchers have built public installations around exactly this, and the results can be striking — not as documentation of movement, but as a real-time translation of it.

It can generate music and soundscapes that respond to you. Audio generation models, coupled with movement tracking, can produce ambient sound environments that shift with your pace, elevation, or stillness. For practitioners who work with sound and vibration, this is one of the more immediately useful current possibilities.

The Gaps That Matter Most to Somatic Practitioners

Here is where honest assessment becomes essential — because the gaps are not minor details. They go to the heart of what somatic practice is.

AI cannot sense effort or quality. There is a profound difference between a shoulder that is held, a shoulder that is released, and a shoulder that is consciously softened by someone who has spent years learning to distinguish the two. Camera-based systems see geometry: where your arm is in space, the angle of your elbow. They cannot detect neuromuscular tone, tissue quality, or the quality of attention behind a movement. The system does not know whether you are present or dissociated.

It cannot perceive breath — reliably or meaningfully. Breath can be inferred roughly from chest expansion in high-quality video under ideal lighting conditions, but the inference is coarse and easily confused. More importantly, the relationship between breath and movement that anchors most somatic practice — breath as a signal of effort, regulation, openness, resistance — is invisible to current systems.

It has no access to the felt sense. The organism's self-experience — what Eugene Gendlin called the felt sense, what Moshe Feldenkrais tracked through learning and function — is precisely what somatic practice cultivates. It is interior. It is not recoverable from outside. This is not a temporary technical gap waiting to be closed by a better sensor. It is a conceptual boundary.

It cannot distinguish habitual from chosen. When a practitioner asks a student to notice whether they always reach with the same arm first, the inquiry is about pattern recognition at the level of lived experience. AI can identify that you reached with your right arm. It cannot tell you whether you knew you were doing it, or whether you had a choice.

Where Experimentation Is Worth Your Time

Despite all of this, there are genuine entry points — if you approach them with the same quality of inquiry you bring to the mat.

Use pose feedback as a mirror, not a judge. Reviewing your movement through a pose-estimation filter can reveal spatial habits you genuinely do not notice in your body — the asymmetry in your standing, the way your head leads in a particular direction. Treat it the way you might treat a mirror in a studio: useful data, not the whole picture, and never a replacement for a teacher who can sense what they see.

Experiment with AI-generated soundscapes as a field condition. Generative audio that responds to your movement creates a feedback loop without prescribing an outcome. This is actually aligned with certain somatic pedagogies — the environment changes with you, without evaluating you.

Use large language models as a reading companion for somatic theory. The literature underlying somatic practice — phenomenology, embodied cognition, developmental movement — is dense. A well-prompted language model can help you navigate Merleau-Ponty, summarise research on interoception, or generate questions for your own inquiry. It cannot replace embodied learning. It can accelerate theoretical orientation.

Approach generative visual art as contemplation, not documentation. If you practice with a pose-to-image tool, notice what the generated images do to your sense of your movement rather than using them to evaluate it. The aestheticisation of your skeleton can be genuinely interesting as a contemplative object — something strange and external that was nevertheless made from you.

What to Watch for in the Next Two to Three Years

The field is moving fast enough that honest forecasting requires humility, but several directions are technically credible.

Multimodal effort estimation. Wearables that measure muscle activation, skin conductance, and micro-movement tremor, fused with video, are beginning to produce richer profiles of effort and regulation. The data is still coarse relative to a practitioner's perception, but the trajectory is clear.

Longer-window temporal modelling. Current systems are relatively good at a moment of movement; they are weaker at tracking the arc of a session — how quality shifts, how patterns emerge and dissolve over twenty minutes of practice. Improving temporal depth is an active area of development.

Practitioner-trained micro-models. Small, fine-tuned models trained on a specific teacher's or practitioner's movement vocabulary — rather than generic motion-capture datasets — may begin to produce feedback that has genuine pedagogical relevance. The ethics and epistemology of this will require careful attention from the somatic community.

The technology is not coming for your practice. But it is arriving in your field, and it will be shaped — partly — by whether people with embodied knowledge engage with it critically, or leave the design entirely to engineers.

Your felt sense is not a gap in the data. It is the thing the data is trying to reach.