Evidence behind the curriculum.

I · Overview

The study tested whether six lessons of the Pixels2GenAI curriculum, drawn from Modules 1, 3, 4, 9 and 12, could produce measurable conceptual change in a single one-day cohort, and whether the cognitive load of doing so stayed within tolerable bounds for novices. Five research questions framed the work; four are addressed empirically, the fifth — touchpoints with real-time systems — is addressed theoretically because the relevant content did not ship in this iteration of the curriculum. The headline finding is large in magnitude and hedged in interpretation: every participant who completed both tests gained ground, and the correlation between prior experience and reported cognitive load was strongly negative.

II · Method

Design-Based Research (Cycle i of three) shaped both the curriculum and its evaluation. Nine participants — recruited through personal and university networks and posters, demographics ranging from 18 to 44 years and from high-school to master’s-level education — completed a structured one-day workshop covering six lessons drawn from Modules 1, 3, 4, 9 and 12 (Pixel Fundamentals, Transformations, Fractals, Neural Networks and Generative AI). A 24-item conceptual instrument was administered before and after the workshop, organised into four six-item sections covering arrays and images (A), computational concepts: cellular automata, convolution, fractals (B), neural networks (C), and generative AI (D). The NASA Task Load Index was administered after every module to sample cognitive load on six dimensions; a five-prompt exit ticket after every module collected open-form reflection. One participant did not complete both test administrations, so the pre/post analysis is reported on n = 8 while NASA-TLX, exit-ticket, and demographic analyses use the full n = 9.

The validation audit (post-collection) re-checked scoring keys, missing-data treatment, and instrument reliability before any inferential test was run. Reliability statistics are summarised below; the full audit and post-validation action plan are reported in the thesis.

III · Framework design

RQ 1 · What pedagogical principles and design patterns effectively scaffold learning progressions from basic array manipulation to generative AI in creative contexts?

The framework that emerged from cycle i treats every lesson as a triangle: a visual artifact the learner produces, the array operation that generates it, and the conceptual claim the operation makes about the domain. Modules sequence by what changes per step — first values inside an array, then geometry across an array, then transformations between arrays, then learned transformations. This sequencing is the visible commitment of the curriculum; the empirical question is whether it scaffolds. The pilot points that way. Hands-on lessons carried lower mental demand than conceptual deep-dives (9.7 vs 12.7 on a 20-point scale, p = .016, d = 1.10), and self-rated understanding was highest for convolution, RGB basics and fractal square (3.9, 3.7 and 3.6 of 5). The knowledge gains in §VI and the exit-ticket themes in §VIII add to this; the experience-load correlation in §IV suggests the scaffolding is calibrated for the novice end of the experience distribution, with implications discussed in §IX.

IV · Cognitive load

RQ 2 · How can complex technical concepts be decomposed and sequenced to maintain optimal cognitive load while building toward advanced applications?

Workload rose and fell in a sawtooth across the day rather than climbing steadily: lowest at RGB basics (7.7 of 20), highest at cellular automata (10.9), easing at convolution (9.9) and fractal square (8.6), then high again for the perceptron and DDPM (10.7 each). Mental demand varied most between modules (p < .001). Temporal demand also varied (p = .002), peaking at cellular automata and DDPM, so time pressure was part of the picture. By the last two modules participants worked harder, felt less successful and grew more frustrated.

Fig. 1 NASA-TLX trajectory across six modules.

Mean overall workload per module (1–20, n = 9) with bootstrap 95% intervals; Friedman χ²(5) = 15.8, p = .008. Shading marks hands-on (blue) and conceptual (orange) lessons.

The more consequential finding sits sideways across this trajectory. A Spearman correlation between a prior-experience composite (programming years, Python proficiency, NumPy proficiency, image-processing exposure, ML basics) and mean NASA-TLX load across the workshop returned a strong negative coefficient — far larger than is usually reported in load research and significant despite the small sample.

Fig. 2 Correlation heat map · experience, gain, load.

Pairwise Spearman correlations across the load-bearing measures. The experience × load cell (top-right cluster) is the dark band in the map.

V · Integration pathways

RQ 3 · What strategies effectively integrate real-time systems with progressive AI learning, and at what points in the curriculum should these integrations occur?

The pilot did not include TouchDesigner content — Modules 10 and 11, where real-time integration would have been evaluated, had not been released at the time of the workshop. RQ 3 is therefore addressed theoretically: the curriculum reserves the integration question for Cycle ii, whose sixth workshop runs in TouchDesigner, and stages a hypothetical placement against three constraints (the learner’s grasp of arrays as data structures, exposure to neural inference, and the conceptual cost of a second runtime). The theoretical argument and proposed empirical follow-up are developed in the thesis. The relevant disclosure here is that this research question, unlike the other four, lacks empirical receipts in cycle i.

VI · Assessment and gain

RQ 4 · How can learning outcomes in creative AI education be assessed across technical proficiency, creative expression, and conceptual understanding?

Across the 24-item conceptual instrument, every participant who completed both administrations improved. The mean score rose from M = 3.6 to M = 12.1 out of 24 — a Cohen’s d of 1.615. The Wilcoxon signed-rank test on paired pre/post scores returned a significant positive shift (W = 0, p = .008, exact); the normalised gain (Hake’s g) averaged 0.43, classifying the cohort as a “medium” gainer in physics-education-research terminology. In a sample of this size these numbers are signals, not certainties — but the unanimity of the direction (no participant regressed; no participant gained zero) is the part that warrants the headline.

Per-section movement tells a sharper story than the aggregate. Arrays and images (A) moved most (d = 1.69); neural networks (C) and computational concepts (B) also improved significantly. Generative AI (D) moved least: from 0.5 to 1.9 of 6, not significant (p = .062), with 56% of answers still “I don’t know”. Most participants struggled to install PyTorch for the DDPM module late in the day, which the thesis treats as a design lesson for the next cycle.

The assessment design itself is part of the answer to RQ 4. The instrument combines a forced-response conceptual measure (the 24 items reported here) with the exit-ticket free-response coding reported in §VIII, on the argument that creative-AI competence is not adequately measured by either alone. The pilot’s evidence supports the joint instrument: the two strands agreed on the main points (learning gains, load rising at the conceptual modules, the weak transfer to generative AI) and are combined in the thesis’s integration matrix.

VII · Transfer

RQ 5 · To what extent do learners successfully transfer foundational computational concepts to novel creative AI contexts, and what factors facilitate this transfer?

Transfer was sampled rather than measured: the pilot was a single day, and follow-up administration was not part of cycle i. The exit-ticket coding nevertheless offers signal on intended re-use. Codes in the prospective-transfer family — participants naming a specific later context where they plan to apply what they learned — appeared in seven of nine exit tickets. The named contexts cluster into three groups: personal art projects (four mentions), immersive-installation work (two), and improving Python skills for an existing professional context (three). The integration matrix in §VIII shows the alignment between these intended re-uses and the conceptual gains in §VI.

VIII · Themes

The exit-ticket coding triangulates the quantitative findings from a register the test items cannot reach. 158 responses were coded with a 30-code codebook in five groups (learning, connections, challenges, engagement, experience), 204 code applications in all. The most frequent codes: naming a concept learned (31), an explicit link to an earlier lesson (16), a conceptual difficulty (12), and setup problems (8, half of them in the DDPM module).

Fig. 4 Exit-ticket theme frequencies.

Code frequencies from the exit tickets (158 coded responses, 30-code codebook). Naming a concept learned is by far the most frequent code.

By the end of the day participants used array, channel and kernel in their own sentences, often without prompting: the qualitative counterpart of the knowledge gains in §VI. The curriculum’s scaffold is Execute → Modify → Create, and the integration matrix in the thesis lines up both strands of evidence theme by theme.

IX · Limitations

The page above is the page; receipts stand behind it.