Parse: Teaching How AI Learns Through Tangible Unplugged Modules

Parse minimizes reliance on prior technical knowledge and digital infrastructure, enabling use in low-resource contexts. Embodied cognition research establishes that physical interaction with manipulable objects supports conceptual understanding of abstract STEM processes [14]. Tangible interfaces have externalized algorithmic processes into physical form, with recent work showing that physically manipulable ML tools prompt reflection about model behavior [6] and that embodied interaction contributes to AI learning outcomes [8]. Within AI education, the "AI Unplugged" framework [7] covers classification, neural networks, and reinforcement learning through card-sorting and role-play, while subsequent work has extended to GAN simulation [12] and kinesthetic exercises for conversational AI [10]. Yet these activities remain concentrated on supervised paradigms. Research shows learners hold persistent misconceptions about how ML systems learn from data [9], suggesting that learning from feedback may require direct physical engagement with pattern discovery processes. The constructionist tradition supports this direction [5].

Parse was developed through a Research through Design process across five sequential workshop cohorts. Each cohort informed revisions prior to the next deployment: observations and performance data were consolidated between sessions and used to evaluate whether each module's physical form and interaction flow adequately externalized its target concept. This iterative cycle is itself a core contribution: the design knowledge produced through successive classroom deployments shapes both the final artifacts and our understanding of what tangible AI learning systems require. Module 1 underwent the most significant redesign. In early deployments, two problems emerged simultaneously: the mechanism proved mechanically fragile, and the interaction resolved too quickly, preventing students from developing the systematic search strategy the module was designed to build. The redesign extended the interaction to a two-stage configuration and simplified the mechanical structure. In subsequent cohorts, observers recorded a clear behavioral shift: students began adopting sequential, exhaustive strategies within two to three attempts. Modules 2 and 3 remained structurally stable across cohorts. This redesign illustrates a broadly applicable constraint: the temporal structure of an interaction — how long it sustains engagement before resolving — may be as pedagogically significant as its conceptual mapping.

We use feedback-derived learning as an umbrella framing for processes in which a system discovers structure without labeled outputs — encompassing random search, iterative narrowing, and distributional inference as mechanistically distinct but conceptually related challenges. Module 1 externalizes random search: exhaustive exploration of a solution space with no proximity signal [1]. Module 2 externalizes iterative constraint propagation: each solved component reduces the remaining search space, mirroring how iterative algorithms narrow candidate solutions [15]. Module 3 externalizes statistical inference from accumulated evidence: only by tracking outcomes across multiple interactions can learners build a model of the hidden distribution. Module 3 was retained across all cohorts precisely because its divergent outcomes reveal a boundary condition: tangible interaction alone is insufficient when feedback must be accumulated rather than immediately interpreted.

Module 1: Random Search. Design principle: No proximity signal forces exhaustive exploration. The module contains two wheels, each with eight orientations: seven presenting a wall and one open. The learner rotates both wheels, locks them, and drops a ball into the entry hole. If both open paths align, the ball passes through; otherwise the learner retrieves it and tries again. The configuration space is 64 combinations, of which only one permits passage. Because the internal structure is hidden, the learner must explore exhaustively. This interaction reflects how learning systems explore a hypothesis space by systematically evaluating solutions in the absence of guiding feedback, analogous to brute-force or grid search strategies in ML.

Module 2: Iterative Refinement. Design principle: Immediate physical feedback accelerates conceptual grounding. The module contains six cylindrical tubes with unique geometric profiles and six matching rectangular disks. Three tubes are placed inside an enclosure with their profiles hidden. A correct match produces an audible click as the tube seats. Each successful match progressively narrows the solution space. This process parallels how learning systems iteratively eliminate inconsistent hypotheses, progressively narrowing the solution space in a manner similar to constraint satisfaction and version space reduction.

Module 1: Random Search, Module 2: Iterative Refinement both CAD visuals through Fusion360 and real-life modules

Module 3: Strategic Inference. Design principle: Delayed, distributed feedback requires active model-building. The module contains seven sliding keys, each with three holes and three notches. Three balls are distributed across a hidden grid. Pure guessing produces slow results; tracking which coordinates have yielded balls across rounds builds a progressively more accurate model of the hidden distribution. This interaction reflects how learning systems construct models by aggregating observations over time, analogous to statistical inference and the estimation of underlying data distributions.

Module 3: Strategic Inference's both CAD model through Fusion360 and real-life model

Five sequential workshops were held with 120 middle school students (grades 5–8), in groups of 24 from the same school. Each session lasted 80 minutes: a 10-minute introduction (without disclosing the specific concepts targeted), three module interactions of approximately 15–18 minutes each, and a 10-minute post-workshop survey. Students worked in small groups of two to three and engaged with the modules sequentially, one at a time. Within each group, participants collaboratively explored the interaction, discussing strategies and observations as they attempted to solve each module. Interaction continued until the group reached a solution or demonstrated a stable approach to the task. For instance, during interpreting Module 1, one group initially rotated both wheels randomly, but after repeated failure, a student within the group suggested fixing one wheel and systematically rotating the other, stating ‘we haven't tried all of these yet.’, marking a shift toward systematic search. Systematic search reflects how ML systems explore a hypothesis space by evaluating candidate solutions.

Two interactive images we captured throughout our workshops and demonstrations with middle school students

Two observers independently recorded field notes targeting: initial approach, strategy shifts, reasonings, peer interaction, and time to completion. No formal inter-rater reliability measure was calculated, as observations were qualitative and intended to contextualize survey findings. Pre/post surveys measured self-reported familiarity on a 5-point Likert scale across three target concepts. The post-workshop survey included four multiple-choice assessing whether students could recognize each module's core mechanism in real-life (see Table 1). Results are reported descriptively; no inferential statistics were applied. Since measures rely primarily on self-reported familiarity and a limited number of transfer items, results should be interpreted as indicative rather than conclusive evidence of conceptual understanding.

Pre/post test responses showed gains across all three target concepts. Module 1 showed the largest shift (median 2.0 to 4.0 on a 5-point Likert scale), Module 2 a moderate gain (2.0 to 3.5), and Module 3 the smallest (2.0 to 3.0). Transfer accuracy followed the same gradient: Module 1 at 82Notes contextualize this pattern. In Module 1, most students shifted from unsystematic to sequential search within two to three attempts. Module 2 produced the most visible engagement: students who accelerated through subsequent attempts, frequently stating that the solution space was shrinking. Module 3 produced the most divergent responses: some students, particularly in grades 7–8, adopted tracking strategies early, while generally all grades persisted with random guessing throughout.

This gradient maps translates a structural difference in feedback type. Modules 1 and 2 provide unambiguous, immediate physical signals. Module 3 requires aggregating outcomes: a fundamentally different cognitive demand. The competitive group format may have chosen rapid guessing over tracking, potentially influencing outcomes independently of the intended feedback, accumulation challenge.

The iterative deployment of Parse produced design knowledge beyond what pre/post surveys captured. The redesign of Module 1 illustrates that the temporal structure of an interaction is as pedagogically significant as its conceptual mapping: a module that resolves too quickly does not give learners sufficient time to construct the target inference. Module 2’s immediate auditory feedback produced reliable engagement and rapid conceptual convergence, requiring no redesign. Module 3 remained the most pedagogically challenging across all deployments, suggesting that this difficulty is intrinsic to the concept the module targets (statistical inference) rather than a solvable design problem at this level of scaffolding. Interaction too fast → added second stage → longer engagement.

Our first initial design of the Module 1.

The convenience sample from a single private school limits generalizability, and between-grade variation introduces developmental differences the study was not designed to disentangle. The iterative design process means not all cohorts interacted with identical module versions; results should be interpreted as directional rather than strictly comparable. Replication in genuinely low-resource contexts would be needed to validate the accessibility claims.