Confidence label: Speculative. This brief proposes a non-contact movement experiment in which an AI has a limited budget for asking clarification questions. The question is whether well-timed requests preserve a performer's agency better than either continuous guessing or indiscriminate pausing. No outcome is claimed. The specific combination is a research proposal; the source scan does not establish that no related experiment has ever been attempted.
The provocation
Give a movement-reactive visual instrument permission to ask at most three questions during a short session. It must decide when its uncertainty warrants spending that budget. The performer may answer, skip, or stop, and a skipped question must not be converted into agreement.
The instrument should remain useful when its budget runs out. It can hold a neutral visual state or continue a clearly labelled direct mapping, but it cannot secretly turn an unanswered intention into a confident generative command. The provocation is to treat clarification as a scarce interaction resource rather than a free correction channel.
The argument
Asking more questions can improve information while damaging the activity being supported. A dancer cannot necessarily interrupt a phrase whenever a classifier wants another label. Conversely, a system that never asks may repeatedly impose an interpretation that the person did not intend.
The design problem is therefore not maximum question frequency. It is the allocation of interruption: which moments merit one, what response is possible, and whether the answer changes anything useful. A clarification budget makes that trade-off visible and prevents a prototype from appearing reliable simply because it continually returns responsibility to its user.
The Lab's November synthesis on productive hesitation develops the conceptual connection. This brief specifies a narrower test: whether uncertainty can help place a small number of interventions better than a timing-only baseline.
Current evidence
Selective-classification research provides a precedent for withholding some predictions rather than answering every input (Geifman & El-Yaniv, 2017). AffectiveVLA provides a current example of confidence gating before humanoid execution (Kuzmin et al., 2026). Neither establishes the proposed benefit of clarification in creative practice.
Micro-phenomenological interviewing provides a separate methodological resource for investigating particular experiences rather than relying only on broad satisfaction scores (Petitmengin, 2006). It does not turn participant reports into objective labels of intention. The experiment below combines these resources as a hypothesis, keeping their evidential roles distinct.
The proposed experiment
Phase 1: establish the mapping and the cost of interruption
Recruit a small, explicitly exploratory group of adult movement practitioners. Begin with a non-contact visual instrument using a camera and a simple, documented movement-to-image mapping. Do not infer emotions, diagnose conditions or ask participants to perform strenuous or unfamiliar actions.
Agree on a few optional command meanings with each participant, such as continuing a visual phrase or resetting it. Keep the direct sensing and the inferred command separate in the record. Let participants practise answering through a button, short phrase or skip control; the comparison must not depend on fluent verbal explanation while moving.
Use this phase to identify a tolerable session duration and question format. Freeze those choices before the comparison. Pilot adjustments are not evidence that a particular gating policy works, and the report should distinguish pilot decisions from later observations.
Phase 2: compare where the same budget is spent
Run three counterbalanced conditions: no clarification with a neutral hold on uncertain inputs; three clarification opportunities placed independently of model uncertainty; and up to three opportunities placed by a predefined uncertainty rule. Keep the maximum budget, interface and visual mapping constant for the latter two conditions.
Match or record total hold time so that a benefit cannot be explained simply by slowing the system. Define the uncertainty rule before testing and do not tune it on the evaluation sessions. Record the question's trigger, the evidence available, the answer or skip, the subsequent output and any participant correction.
The random-placement condition is essential. Without it, a positive result could show only that people like occasional check-ins. The experiment asks whether the model chooses useful moments, not whether clarification is pleasant in general.
Phase 3: inspect episodes before revealing model labels
After each condition, invite accounts of selected moments, including at least one question, one uninterrupted passage and one disagreement if present. Begin without displaying the model's interpretation. Only afterwards compare the participant's description with the sensing and decision record.
Measure interruption burden, skipped requests, incorrect command activations, accepted-input coverage and recovery after a correction. Keep qualitative accounts alongside those counts instead of collapsing them into one agency score. A participant may prefer a less accurate but less intrusive condition, and that trade-off should remain visible.
What success looks like
The hypothesis would receive preliminary support if uncertainty-placed questions reduce consequential misinterpretations relative to timing-only questions without raising reported interruption burden or excluding substantial portions of a person's movement. Report session-level results and uncertainty; a small exploratory sample cannot justify a universal threshold.
Failure includes equally important outcomes: random questions work as well, the system asks too late, participants feel pressured to answer, or the neutral-hold baseline is preferable. A useful study may conclude that intention inference should be removed in favour of a predictable direct mapping.
There is also a governance criterion. Every question must have a traceable effect or an explicit reason why it did not change the output. Asking for input while ignoring it is not shared control. The evidence-ledger guide supplies a starting point for documenting that chain.
The invitation
The first implementation needs no humanoid robot and no claim to read inner intention. It needs a stable visual mapping, an inspectable gate, a genuine skip control and a research team willing to preserve disagreements in its report. Consent to participate, consent to recording and consent to training reuse must remain separate choices.
The invitation is to test a small form of restraint. If a system can only ask a few times, can it learn where its interpretation most needs the person's contribution—and can it accept that the person sometimes does not want to resolve the ambiguity?
References
- Geifman, Y., & El-Yaniv, R. (2017). Selective classification for deep neural networks. arXiv. https://arxiv.org/abs/1705.08500
- Kuzmin, N., Demianchuk, G., Lezina, M., Mahmoud, Y., Zhura, I., Altamirano Cabrera, M., & Tsetserukou, D. (2026). AffectiveVLA: Safe and intent-aware VLA control for human-humanoid social interaction. arXiv. https://arxiv.org/abs/2611.07712
- Petitmengin, C. (2006). Describing one's subjective experience in the second person: An interview method for the science of consciousness. Phenomenology and the Cognitive Sciences, 5, 229–269. https://doi.org/10.1007/s11097-006-9022-2