Confidence label: Plausible. Micro-phenomenological inquiry and selective prediction share a productive problem: how to avoid turning insufficient evidence into a premature answer. They do not share a mechanism or a theory of experience. The connection proposed here is methodological: a movement-reactive AI system can make room for clarification without pretending that model uncertainty measures the moving person's experience, or that a confident classification grants permission to act.
The scope is human–AI co-creation rather than autonomous physical contact. Foundational first-person-method literature supplies the inquiry side; selective-classification research and a November 2026 intent-gated robotics preprint supply the computational side. The proposal is a design hypothesis, not a demonstrated benefit of combining those literatures.
The shared problem is premature closure
A performer makes a small movement before a larger phrase. A classifier assigns it a label; the visual system responds immediately. The interface may feel responsive, yet the label could have closed the interaction too early. Perhaps the movement was preparatory, perhaps it was exploratory, or perhaps the camera simply lost the hand. These alternatives call for different responses, not merely a faster prediction.
The same problem appears in research conversation. A participant is asked what a movement felt like and quickly offers a familiar explanation. That explanation may be meaningful without providing a fine-grained account of a particular event. A method that seeks detail needs a way to return to the event instead of treating the first available concept as the whole description.
Neither situation is solved by refusing every answer. Permanent hesitation prevents a system from participating and an interview from producing a usable account. The useful question is when to proceed, when to ask and how to preserve what remains unknown. That is the shared structure explored here; it is not evidence that computational rejection and experiential inquiry are equivalent.
What the two literatures actually establish
Petitmengin (2006) presents an interview method for eliciting detailed accounts of subjective experience and examines difficulties of awareness, description and validity. Prpa and colleagues (2020) investigate micro-phenomenology in interaction-design research through interviews with experienced practitioners of the method. Their work supports taking the elicitation process seriously rather than treating immediate verbal response as transparent access to experience. It does not make an interviewer an infallible reader of another person's body.
On the computational side, Geifman and El-Yaniv (2017) formulate selective classification as a trade-off between error on accepted predictions and the proportion of inputs a classifier accepts. Rejecting some cases can improve the reliability of those retained under the method's assumptions. That result concerns tested classification settings, not an automatic guarantee for an unfamiliar dance studio.
AffectiveVLA adds a current embodied example. Kuzmin and colleagues (2026) place an intent-confidence gate before motion execution in a humanoid interaction pipeline. This is evidence that withholding can be an explicit architectural operation. It is not evidence that confidence is consent or that the entire system is certified safe. The synthesis uses that narrow architectural fact, not the paper's title, as its point of contact.
A pause needs a reason and a next step
The Lab proposes distinguishing four reasons for withholding. Signal uncertainty means the sensing record is weak. Model uncertainty means the system cannot reliably choose among its available interpretations. Semantic ambiguity means several interpretations may be reasonable even with good sensing. Permission uncertainty means the person has not authorised the response. These categories are a design vocabulary proposed here, not a taxonomy validated by the cited studies.
Their distinction changes the next action. Weak sensing might justify asking the participant to move back into view. Uncertain classification might justify a neutral output while awaiting more evidence. Ambiguous meaning might justify offering alternatives. Missing permission requires asking or remaining inactive, even when the classifier is certain.
A single confidence threshold cannot perform all four jobs. If a high-confidence system automatically proceeds, it may conceal semantic disagreement or ignore a person's decision. If every uncertainty triggers the same interruption, it can make participation exhausting. A useful pause therefore has to name its reason and provide a proportionate route forward.
Why the connection is more than an analogy
The proposal changes an interface contract. Instead of requiring one movement label per frame, the system can return a structured state: what was observed, what was inferred, whether it is withholding, why, and what additional evidence or decision would permit continuation. The output may include an unanswered question without treating that question as a technical error.
The corresponding research protocol also changes. A participant's initial interpretation need not be overwritten when they later describe the event differently. Both can remain in the record, with the elicitation context attached. This preserves a distinction between a quick label and a later account without assuming that later is always truer.
The archive's evidence-ledger guide already separates observation, inference and generation. Productive hesitation adds a decision layer: why the system did not convert its current inference into action. Its value would have to be tested, but its data requirements are concrete enough to implement and inspect.
What movement inquiry could gain
A computational system can timestamp the moment a tentative label was formed, the evidence it used and the response it withheld. Such a record could help an interviewer locate a specific episode rather than asking only for a general impression of the whole session. It would also expose occasions when the system's uncertainty had nothing to do with the participant's felt ambiguity.
This should remain a supporting record, not an interview script that dictates experience. Asking “the model detected hesitation; where did you feel it?” would already import the model's category. A less leading route would begin with the participant's account of the episode and compare it with system records afterwards. The proposed order is an experimental safeguard, not a claim that one phrasing guarantees an uncontaminated report.
Movement researchers could also distinguish a useful pause from a frustrating delay. A participant may welcome a request for clarification at one point and experience the same request as an interruption elsewhere. Logging the sequence permits analysis of those differences without reducing every pause to a latency statistic.
What selective systems could gain
Selective prediction usually needs a criterion for what counts as an error and a decision about acceptable coverage. Co-creative movement makes those criteria harder: an unexpected response may be generative, while a highly accurate label may narrow an exploration the person wanted to keep open. First-person inquiry can help identify which consequences matter to participants without claiming to replace performance measurements.
For example, a participant may distinguish being misunderstood from being productively surprised. The system could have the same classification error in both cases, yet the interaction differs. An evaluation that records only label accuracy would miss that distinction. Conversely, a participant's enjoyment should not excuse unreliable sensing or erase a clear failure to respect a stop request.
This is why the connection is generative. It asks the technical field to broaden the consequences it measures and the experiential field to retain traceable system decisions. Neither is asked to abandon its own standards. The Lab's dance-led research profile provides a related account of practice changing research questions rather than merely supplying examples.
A concrete co-creation study
Consider a non-contact visual instrument in which arm and torso motion shape an abstract image. Compare three conditions: continuous response, confidence-gated neutral holding, and the same gate with a brief optional clarification. Keep the underlying visual mapping and sensing stack fixed, and counterbalance condition order. These are proposed study conditions, not an existing result.
Include a baseline with the same total pause time distributed independently of uncertainty. Otherwise, any benefit could come from slowing the encounter rather than locating useful moments. Record accepted-input coverage, visible mapping failures, interruption count, response delay and participant withdrawal. Then elicit descriptions of selected episodes before showing the model's labels.
The primary question is not whether asking questions is always better. It is whether reason-specific withholding reduces misinterpretation while preserving room for exploration. A null result would be informative if participants find the clarification distracting or if a neutral hold works just as well. The design should be willing to conclude that a simpler interface is preferable.
Important limits and counterarguments
The strongest counterargument is that this proposal overcomplicates a creative instrument. Many useful instruments respond predictably without interpreting the performer's intention at all. If a direct mapping answers the artistic question, adding a classifier and a clarification dialogue may introduce a problem that did not exist. Productive hesitation is relevant when interpretation is already part of the system, not a reason to insert interpretation everywhere.
Another limit is accessibility. Verbal clarification may favour participants comfortable with the language, interviewer or research setting. An interface should permit nonverbal alternatives, a simple skip and withdrawal. Silence must not become evidence of assent or a missing value automatically filled by the model.
Finally, confidence can be poorly calibrated under unfamiliar bodies, clothing, viewpoints or movement styles. A gate can then withhold disproportionately from the people least represented in training. Aggregate error reduction may conceal unequal access to the instrument. Coverage and interruption therefore need to be reported by relevant session conditions, with care not to turn small samples into unsupported demographic claims.
A more accountable form of responsiveness
Responsiveness is often described as immediate output. The connection proposed here suggests a richer criterion: a system responds appropriately to the limits of its evidence and to the person's authority over participation. Sometimes that means continuing; sometimes it means holding, asking or accepting that the intended meaning remains unresolved.
That is not artificial introspection. The system is not acquiring lived uncertainty by emitting an abstention flag. It is providing an inspectable boundary between an inference and an action. Micro-phenomenological inquiry helps formulate questions about that boundary; selective prediction makes one part of it operational. Whether the resulting interaction is better remains a question for careful, practice-led experimentation.
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
- Prpa, M., Fdili Alaoui, S., Schiphorst, T., & Pasquier, P. (2020). Articulating experience: Reflections from experts applying micro-phenomenology to design research in HCI. In Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. https://doi.org/10.1145/3313831.3376664