Evaluating AI-Generated Movement: A Somatic Practitioner's Assessment Framework
How to bring your trained perception to bear on machine-generated movement — the third evaluation axis that the technical field cannot yet measure
Who This Is For
This guide is for movement practitioners — dancers, somatic educators, choreographers, movement therapists — who are being asked to evaluate, give feedback on, or collaborate with AI-generated movement, and who want a structured way to apply their trained perception to the task.
It is also for practitioners who want to understand what they uniquely contribute to somatic-AI work. As the June content on movement evaluation discussed, the technical field can now measure two things about generated movement: whether it is physically valid (obeys the laws of physics) and whether it looks convincing to a general observer. What the field cannot yet measure is the third axis — whether a movement has interior fidelity, whether it is true to the felt experience it expresses. That axis requires trained somatic perception. This guide is a framework for applying it systematically.
Why Your Assessment Is Not Redundant
Before the framework, the premise. When an AI generates movement, why does a somatic practitioner's evaluation add anything that automated metrics and general audiences don't already provide?
Because you perceive dimensions of movement that neither can access. A physics metric checks mechanical validity; it is blind to quality. A general audience perceives gross plausibility ("that looks like a real person") but not fine quality ("that movement is initiated from the periphery and lacks core support"). Your training has cultivated reliable perception of exactly the dimensions that fall between these — the qualitative, organisational, felt dimensions of movement that constitute its expressive truth.
This is not a subjective opinion you happen to hold. It is a trained perceptual capacity, as real and reliable as a musician's trained ear. The framework below is a way of deploying it systematically rather than impressionistically.
The Framework: Five Assessment Dimensions
For any AI-generated movement, assess these five dimensions. Each is something your training lets you perceive and that automated metrics currently cannot.
1. Initiation — Where does the movement begin?
Watch where each movement originates in the body. Does it initiate from the centre and sequence outward (core-initiated), or does it start at the extremities and drag the body behind (peripheral-initiated)? Is the initiation consistent with the movement's apparent intention, or mismatched?
What to look for in AI movement: Generated movement frequently has ambiguous or absent initiation — the body parts arrive at their positions without a clear origin of the movement impulse. This is a hallmark of movement generated from position data rather than from an organising intention. Note where initiation is legible and where it is missing.
2. Sequencing and connectivity — How does movement travel through the body?
In organically organised movement, impulse travels through the body in coherent sequences — the movement of one part relates to and flows into the movement of adjacent parts. Assess whether the generated movement has genuine through-body connectivity or whether body segments move in parallel without relationship.
What to look for: Generated movement often shows segmental independence — limbs that move correctly but without felt connection to the torso or to each other. The joints hit their targets, but the body doesn't move as an integrated whole. Note where connectivity is present and where the body fragments.
3. Effort quality — What are the Laban Effort characteristics?
Apply the Effort framework (Weight, Time, Space, Flow). Does the movement have a coherent, legible effort quality, or is the effort quality flat, generic, or internally inconsistent? Does the effort match the movement's apparent intention?
What to look for: Generated movement tends toward effort-neutrality — a middle-register quality that is neither clearly strong nor light, neither bound nor free. Where the movement should have a distinct effort character, note whether it does. Flat effort is one of the clearest signatures of movement generated without a quality-level model. (See the May 28 practitioner guide for the full Effort vocabulary.)
4. Breath and phrasing — Does the movement breathe?
Organic movement has phrasing — moments of gathering and release, of preparation and completion, that correspond to the natural rhythm of breath and effort. Assess whether the generated movement has coherent phrasing or whether it is metrically uniform, without the breath-like ebb and flow of lived movement.
What to look for: Generated movement is often unphrased — technically continuous but without the felt punctuation of preparation, arrival, and recovery. It moves at a uniform "temperature." Note where phrasing emerges and where the movement runs flat.
5. Intention legibility — Can you read what the movement is for?
The highest-level question. Watching the movement, can you perceive an organising intention — a sense that the movement is about something, reaching toward something, expressing something? Or does it read as well-formed but hollow, correct in form but without a legible reason for being?
What to look for: This is where generated movement most often falls short, and where your perception is most valuable. Movement can pass every technical check and still read as intentionless. Name specifically what intention you can or cannot read.
How to Deliver the Assessment
For each of the five dimensions, provide:
- A specific observation — not "it looks unnatural" but "the reach in the second phrase has no clear initiation; the arm arrives at its endpoint without an originating impulse from the core."
- The felt/qualitative consequence — what this does to the movement's quality and truth.
- A directional note — what would move it toward greater fidelity, in somatic terms.
Specificity is what makes your assessment usable. "It feels lifeless" is true but not actionable. "The movement lacks core-initiated sequencing and reads as segmentally independent, which is why it feels mechanical rather than alive" gives the technical collaborators something they can work with — even if they cannot yet perceive it themselves.
Working With the Technical Team
A note on the collaboration itself. When you deliver an assessment along these five dimensions, you will often be describing qualities your technical collaborators cannot directly perceive. This is not a communication failure — it is the entire point. You are providing perceptual access to dimensions their tools and training do not reach.
Two practices help. First, where possible, show rather than only tell — demonstrate the difference between the generated movement and a somatically fuller version with your own body, so the quality becomes visible even to an untrained eye through direct contrast. Second, connect your observations to the signals the team can measure where a link exists — e.g. flat effort quality often corresponds to absent variation in the dynamic/acceleration profile; segmental independence to a lack of correlated timing across joints. Building these bridges over time helps the felt dimension gradually become partially legible to the technical process.
The Larger Point
This framework operationalises something worth stating plainly: the somatic practitioner is not a consumer of AI movement tools but a bearer of perceptual knowledge the tools do not contain. The five-dimension assessment is a way of making that knowledge explicit, systematic, and usable — of turning trained perception into actionable evaluation.
As the field's automated metrics improve, the physically-valid and generally-convincing axes will increasingly be handled by machines. The interior-fidelity axis will remain, for the foreseeable future, the domain of trained human perception. This framework is a way of occupying that domain deliberately and rigorously.
Further Reading
Zhao, S., et al. (2025). PP-Motion: Physical-perceptual fidelity evaluation for human motion generation. arXiv:2508.08179. https://arxiv.org/abs/2508.08179
Newlove, J., & Dalby, J. (2004). Laban for actors and dancers. Nick Hern Books.
Schiphorst, T. (2011). Self-evidence: Applying somatic connoisseurship to experience design. CHI '11 Extended Abstracts. https://doi.org/10.1145/1979742.1979640