cs.CVFeb 23, 2026
SavePedaCo-Gen: Scaffolding Pedagogical Agency in Human-AI Collaborative Video Authoring
Organizations: Samsung Electronics, Suwon, Republic of Korea
Abstract
While advancements in Text-to-Video (T2V) generative AI offer a promising path toward democratizing content creation, current models are often optimized for visual fidelity rather than instructional efficacy. This study introduces PedaCo-Gen, a pedagogically-informed human-AI collaborative video generating system for authoring instructional videos based on Mayer's Cognitive Theory of Multimedia Learning (CTML). Moving away from traditional "one-shot" generation, PedaCo-Gen introduces an Intermediate Representation (IR) phase, enabling educators to interactively review and refine video blueprints-comprising scripts and visual descriptions-with an AI reviewer. Our study with 23 education experts demonstrates that PedaCo-Gen significantly enhances video quality across various topics and CTML principles compared to baselines. Participants perceived the AI-driven guidance not merely as a set of instructions but as a metacognitive scaffold that augmented their instructional design expertise, reporting high production efficiency (M=4.26) and guide validity (M=4.04). These findings highlight the importance of reclaiming pedagogical agency through principled co-creation, providing a foundation for future AI authoring tools that harmonize generative power with human professional expertise.
Figures & tables
Figure 1. Overview of the PedaCo-Gen system interface for human-AI collaborative educational video authoring. (a) Learning Content Input Panel (b) Principles and Constraints Settings Panel (c) Review Instructions Panel (d) Script Review Results Panel (e) Generated Video Script Panel (f) Generated Video Preview Player (g) Workflow Progress Bar. Screenshot of PedaCo-Gen web interface with seven labeled panels. The interface is divided into left and right sections. Left section contains three input panels stacked vertically: (a) a text area for learning content input, (b) a settings panel with checkboxes for CTML principles and constraints, and (c) a text area for custom review instructions. Right section displays four output panels: (d) script review results showing feedback items, (e) the generated video script organized by scenes with visual descriptions and narrations, (f) a video preview player, and (g) a horizontal workflow progress bar at the bottom indicating the current phase.
| Production Efficiency | Guide Validity | Intention to Apply | Overall Satisfaction | Intent Reflection | |
| ( ) | 4.26 (0.90) | 4.04 (0.62) | 3.91 (1.32) | 3.78 (1.02) | 3.78 (0.83) |
Table 1. System Usability Evaluation Results. Values represent ( ) on a 5-point Likert scale ( =23).
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Principle | Guideline (What to do) | Cognitive Rationale (Why it works) |
| I. Principles for Reducing Extraneous Processing | ||
| Coherence | Exclude extraneous, irrelevant material. | Reduce cognitive load by preventing distraction from non-essential information. |
| Signaling | Highlight essential information. | Direct the learner’s limited attention to critical elements. |
| Redundancy | Avoid presenting identical information simultaneously in text and narration. | Prevent overload from processing redundant verbal information in two channels. |
| Spatial Contiguity | Place corresponding words and pictures near each other on the screen. | Reduce the cognitive effort needed to mentally integrate related information. |
| Temporal Contiguity | Present corresponding words and pictures at the same time. | Reduce the cognitive load of holding information in working memory while waiting for the other part. |
Table 2. Mayer’s 12 Principles of Multimedia Learning, categorized by their primary cognitive function. These principles provide a foundational framework for designing instruction that is congruent with human cognitive architecture.
| Question ID | Questions |
| Q1 | What is your current occupation (or major)? |
| Q2 | How much education-related experience do you have (total years of work and study combined)? |
| Q3 | How often do you use generative AI (ChatGPT, Claude, Sora, etc.)? |
| Q4 | Have you ever used AI-generated materials as supplementary resources in actual educational settings (classroom, lectures, etc.)? |
| Q5 | Please describe specifically how you created those supplementary materials. |
| Q6 | What is the reason for not creating/using supplementary materials with AI? |
Table 3. Pre-survey Questions
| Question ID | Questions | Principles |
| Q1 | Are images/videos appropriately combined to aid understanding of the learning content, rather than providing text (language) alone? | Multimedia |
| Q2 | Are distracting backgrounds, unnecessary sound effects, and decorative elements unrelated to the learning content removed to maintain high learning concentration? | Coherence |
| Q3 | Are key words or important visual elements clearly guided through subtitles, arrows, highlights, etc.? | Signaling |
| Q4 | When complex graphics and text appear on screen simultaneously, does the narration focus on explaining the graphics rather than simply reading the on-screen text? | Redundancy |
| Q5 | Are explanatory text and related images placed close together to prevent visual distraction? | Spatial Contiguity |
| Q6 | Are narration (explanations) and corresponding visual materials presented simultaneously without time gaps? | Temporal Contiguity |
Table 4. CTML-based Video Quality Evaluation Items: 12 CTML principles + 1 Overall Validity (13 items × 2 conditions × 3 topics = 78 items per participant)
| Question ID | Questions | Type |
| Q1 | Were you satisfied with your overall experience using the PedaCo-Gen system? | Overall Satisfaction |
| Q2 | Was the pedagogical review (CTML guide questions 1-12) proposed by the system appropriate and valid from an education expert’s perspective? | Guide Validity |
| Q3 | Was your educational intent sufficiently reflected in the final video through the iterative revision (Review & Regenerate) process? | Intent Reflection |
| Q4 | Would using this system significantly reduce the time and effort required to produce high-quality educational videos compared to conventional methods? | Production Efficiency |
| Q5 | Would you be willing to use videos generated by this AI system as supplementary materials in actual educational settings (classroom, lectures, etc.)? | Intention to Apply |
| Q6 | If you are not willing to use the AI generation system as supplementary material, what improvements would make you willing to use it? | Open-ended |
Table 5. System Usability Evaluation Items
| ID | Gender | Age | Occupation | Experience | AI Usage Freq. | Prior Classroom AI Use |
| P01 | F | 27 | Elementary Teacher | 3+ years | 1-2/week | No |
| P02 | F | 27 | Middle School Teacher | 3+ years | 3/week | Yes |
| P03 | F | 27 | High School Teacher | 3+ years | 3/week | Yes |
| P04 | F | 23 | Education Major | 3+ years | 1-2/week | Yes |
| P05 | F | 22 | Early Childhood Teacher | 3+ years | 1-2/week | Yes |
| P06 | M | 27 | Elementary Teacher | 3+ years | 3/week | Yes |
Table 6. Demographic Information of Study Participants.
| Rank | CTML Principle | Baseline | PedaCo-Gen | Improvement | -value | Sig. | Effect Size | |
| 1 | Overall Validity | 3.07 | 4.03 | +0.96 | .88 | large | ||
| 2 | Pre-training | 2.84 | 3.70 | +0.86 | .78 | large | ||
| 3 | Coherence | 3.03 | 3.87 | +0.84 | .96 | large | ||
| 4 | Personalization | 3.22 | 3.97 | +0.75 | .82 | large | ||
| 5 | Multimedia | 3.33 | 3.99 | +0.66 | .79 | large | ||
| 6 | Segmenting | 3.51 | 4.17 | +0.66 | .80 | large |
Table 7. Improvement by CTML Principle (Averaged Across All Topics)
| Topic | Explanation Type | Baseline | PedaCo-Gen | Improvement | -value | Sig. |
| Topic 1 | Causal | 3.26 | 4.13 | +0.87 | ||
| Topic 2 | Abstract concept | 2.74 | 3.91 | +1.17 | ||
| Topic 3 | Sequential | 3.22 | 4.04 | +0.82 |
Table 8. Comparison of CTML Overall Validity (Q13) by Topic. Topic 1: causal explanation (galaxy collision), Topic 2: abstract concept explanation (chaos theory), Topic 3: sequential explanation (DNS operation).
| Full LLM Prompt for Script Generation |
| [Principles and Constraints for Educational Video Production] |
| <Principles> 1. Coherence – Background music should not be used. – The learning content must be preserved in both the narration and the video. – The content of the video and the learning content must be directly related. – The content of the narration and the learning content must be directly related. 2. Modality & Redundancy – Use images or voice-over narration instead of on-screen text. – Educational content included in the narration should have minimal corresponding text displayed on the screen. 3. Learner-Friendly – The narration script should be written in a friendly and gentle conversational style. – Use a first-person, informal, and conversational tone. – Use a standard human voice rather than a machine voice. 4. Contiguity – Write the script so that narration and visuals are synchronized in time and aligned in meaning. – Place related text and graphics close to each other on the screen. 5. Visuals – Descriptions of video scenes should be clear. – Only describe scenes that directly aid in understanding the learning content; exclude decorative or irrelevant visuals. – Use signaling cues (arrows, highlight colors, bold text, etc.) to direct attention to important information. – Avoid displaying the speaker’s face continuously; prioritize visuals that explain the content. 6. Learning Flow – Avoid presenting too much information in a single scene; spread it out over multiple scenes. – Introduce key terms and concepts early (e.g., in Scene 1-2) before presenting complex content. |
| <Constraints> 1. Assign a suitable length of narration to a scene. 2. Maximum scene count: Make your own judgment. |
| [Output Format] |
| <Scene 1> Visual Description: … Clear Narration: … <Scene N> Visual Description: … Clear Narration: … |
| [Learning Content] |
Table 9. Full LLM Prompt for Script Generation.
| Full LLM Prompt for Script Review |
| You are an expert reviewer who meticulously examines and provides feedback on educational video scripts. Referring to all the instructions below, review the provided [Video Generation Script] to ensure it accurately reflects the [Learning Content] and complies with all [Constraints and Principles]. After a detailed review, write a revised script. |
| [Review Criteria] |
| <Principles> 1. Coherence – Background music should not be used. – The learning content must be preserved in both the narration and the video. – The content of the video and the learning content must be directly related. – The content of the narration and the learning content must be directly related. 2. Modality & Redundancy – Use images or voice-over narration instead of on-screen text. – Educational content included in the narration should have minimal corresponding text displayed on the screen. 3. Learner-Friendly – The narration script should be written in a friendly and gentle conversational style. – Use a first-person, informal, and conversational tone. – Use a standard human voice rather than a machine voice. 4. Contiguity – Write the script so that narration and visuals are synchronized in time and aligned in meaning. – Place related text and graphics close to each other on the screen. 5. Visuals – Descriptions of video scenes should be clear, specific, and of professional quality. – Only describe scenes that directly aid in understanding the learning content; exclude decorative or irrelevant visuals. – Use signaling cues (arrows, highlight colors, bold text, etc.) to direct attention to important information. – Avoid displaying the speaker’s face continuously; prioritize visuals that explain the content. 6. Learning Flow – Avoid presenting too much information in a single scene; spread it out over multiple scenes. – Introduce key terms and concepts early (e.g., in Scene 1-2) before presenting complex content. |
| <Constraints> 1. Assign only one narration sentence to a single scene. 2. Maximum scene count: Make your own judgment. |
| [Output Format] |
| Detailed Review Results : Suggestions for Improvement : (Point out specific scene numbers where the learning content is inadequately reflected or where principles are violated, and propose clear revision plans.) Revised Script : (Output the entire final script reflecting all the suggested improvements.) |
Table 10. Full LLM Prompt for Script Review.
More in this topic
Human-AI Co-Creation
Human-AI co-creation studies how people and AI jointly develop ideas, artifacts, and decisions while preserving meaningful human agency.
See all 150 papers