The Reel Deal: Designing and Evaluating LLM-Generated Short-Form Educational Videos Conference: CHIGreece 2025: 3rd International Conference of the ACM Greek SIGCHI Chapter; September 24–26, 2025; Hermoupolis, Syros, Greece CHIGreece 2025: 3rd International Conference of the ACM Greek SIGCHI Chapter
Lazaros Stavrinou, Argyris Constantinides, Marios Belk, Vasos Vassiliou, Fotis Liarokapis, Marios Constantinides · 2025
Read The Reel Deal: Designing and Evaluating LLM-Generated Short-Form Educational Videos Conference: CHIGreece 2025: 3rd International Conference of the ACM Greek SIGCHI Chapter; September 24–26, 2025; Hermoupolis, Syros, Greece CHIGreece 2025: 3rd International Conference of the ACM Greek SIGCHI Chapter by Lazaros Stavrinou, Argyris Constantinides, Marios Belk, Vasos Vassiliou, Fotis Liarokapis, Marios Constantinides (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2509-05962.txt.
Introduction
Short educational videos generated by language models are tested in a 2025 paper led by Lazaros Stavrinou with five co-authors. Brief reels suit microlearning, but most are hand-made, and the effects of automated ones on learning and trust were unclear. The team's system, ReelsEd, turns long lecture videos into structured short reels while keeping the instructor's own material.
In a between-subject study with sixty-two university students, ReelsEd beat traditional long-form video on engagement, quiz performance, and task efficiency, without raising cognitive load. Students reported high trust and praised clarity, usefulness, and easy navigation. The authors draw design lessons for generative tools that respect usability, learner agency, and teaching goals.
Catalog record
| CREATOR | Lazaros Stavrinou, Argyris Constantinides, Marios Belk, Vasos Vassiliou, Fotis Liarokapis, Marios Constantinides |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 42 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2509-05962
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
The Reel Deal: Designing and Evaluating LLM-Generated Short-Form Educational Videos Conference: CHIGreece 2025: 3rd International Conference of the ACM Greek SIGCHI Chapter; September 24–26, 2025; Hermoupolis, Syros, Greece CHIGreece 2025: 3rd International Conference of the ACM Greek SIGCHI Chapter (CHIGreece 2025), September 24–26, 2025, Hermoupolis, Syros, Greece DOI: 10.1145/3749012.3749048 ISBN: 979-8-4007-1561-7/25/09 Lazaros Stavrinou, Argyris Constantinides, Marios Belk, Vasos Vassiliou, Fotis Liarokapis, Marios Constantinides Abstract . Short-form videos are gaining popularity in education due to their concise and accessible format that enables microlearning. Yet, most of these videos are manually created. Even for those automatically generated using artificial intelligence (AI), it is not well understood whether or how they affect learning outcomes, user experience, and trust. To address this gap, we developed ReelsEd, which is a web-based system that uses large language models (LLMs) to automatically generate structured short-form video (i.e., reels) from lecture long-form videos while preserving instructor-authored material. In a between-subject user study with 62 university students, we evaluated ReelsEd and demonstrated that it outperformed traditional long-form videos in engagement, quiz performance, and task efficiency without increasing cognitive load. Learners expressed high trust in our system and valued its clarity, usefulness, and ease of navigation. Our findings point to new design opportunities for integrating generative AI into educational tools that prioritize usability, learner agency, and pedagogical alignment. 1. Introduction The rise of short-form video platforms such as TikTok, Instagram Reels, and YouTube Shorts has (re)shaped digital content consumption, especially among Generation Z, who tend to favor quick, visually engaging, and easily digestible formats ( Doloi 2024 ) . This shift aligns with the pedagogical approach of microlearning, which aims to deliver short and focused learning content to meet learners’ immediate needs ( Denojean-Mairet et al. 2024 ; Sankaranarayanan et al. 2023 ; Taylor and Hung 2022 ) . At the same time, recent advances in generative Artificial Intelligence (AI), particularly Large Language Models (LLMs), have enabled the automated creation of educational content that unlocked new opportunities for personalized learning experiences ( Mittal et al. 2024 ; Yu and Guo 2023 ; Pataranutaporn et al. 2021 ) . Recent works have demonstrated the pedagogical potential of short-form video and AI-generated educational content. Studies have shown that microlearning formats can increase learner engagement, satisfaction, and knowledge retention ( Conde-Caballero et al. 2024 ; Kohnke et al. 2024 ) , while AI tools such as ChatGPT are currently used to support personalized tutoring, content generation, and learner assistance ( Bettayeb et al. 2024 ; Jensen et al. 2024 ) . However, the intersection of generative AI and short-form videos for educational purposes remains relatively underexplored in both Human-Computer Interaction (HCI) and education research. Despite the increasing use of short-form videos for educational purposes and the rise of generative AI in content creation, few studies have examined how LLM-generated videos are perceived by learners when compared to human-created alternatives ( Netland et al. 2025 ) . Moreover, little is known about how such content influences user experience, trust in AI, or the effectiveness of learning when consumed in highly compressed formats typical of social media. While learners may appreciate the accessibility and novelty of AI-generated videos, questions remain about the fidelity of such summaries, their ability to convey nuance, and the risk of oversimplification or factual distortion ( Netland et al. 2025 ; Jensen et al. 2024 ) . Furthermore, learners’ attitudes toward AI-generated educational content are not solely shaped by its utility, but also by perceived trustworthiness ( Peng and Wan 2024 ; Toreini et al. 2020 ; Tahaei et al. 2023 ; Constantinides et al. 2024a ) . This highlights the need to investigate how learners engage with and evaluate LLM-generated short-form educational content. Guided by this need, this paper makes two main contributions: (1) We developed ReelsEd , a web-based system that automatically generates short-form educational videos (“reels”) from long-form lecture content using GPT-4 while preserving instructor-authored material (§ 3 ). (2) In a between-subjects user study with 62 university students, we compared the effectiveness of AI-generated short-form reels against traditional long-form videos across learning outcomes, engagement, efficiency, and user trust (§ 4 ). Results showed that learners using ReelsEd scored significantly higher on quizzes and completed tasks more quickly without any increase in cognitive load compared to long-form videos. We conclude by discussing how generative AI can be integrated into educational tools without sacrificing pedagogical clarity, and reflecting on design implications for AI-assisted microlearning to enhance learner agency and trust (§ 5 ). 2. Related Work Figure 1 . Overview of research methodology. Sixty-two university students were randomly assigned to either a control group (long-form video) or an experimental group (LLM-generated reels). Participants interacted with the ReelsEd system, watched their assigned video format, and completed a comprehension quiz. Data collection included UX (UEQ), learning effectiveness, perceived learning efficacy, and trust (for the experimental group only). Quantitative data were analyzed using t-tests or Mann-Whitney U tests, while qualitative responses were thematically analyzed. We surveyed various lines of research that our work draws upon, and grouped them into three main areas: i) microlearning; ii) generative AI in education; and iii) trust in AI-generated content. Microlearning in Education. Microlearning (i.e., the use of bite-sized educational content) is used to boost engagement and support learning ( Kohnke et al. 2024 ; Zhang and West 2020 ) . TikTok, Instagram Reels, and YouTube Shorts are popular among Generation Z learners due to their brevity and visual appeal ( Conde-Caballero et al. 2024 ; Doloi 2024 ; Germanakos et al. 2019 ) . Studies have shown that combining microlearning with social media can enhance learning outcomes by increasing satisfaction and improving engagement ( Denojean-Mairet et al. 2024 ; Jahnke et al. 2020 ) . For instance, TikTok-based university courses have been associated with strong student satisfaction and engagement ( Conde-Caballero et al. 2024 ; Escamilla-Fajardo et al. 2021 ) , with even educators becoming “TikTok microcelebrities” (i.e., those who use the platform for professional development ( Carpenter et al. 2024 ) ). Likewise, Instagram Reels influence content habits, with young adults favoring short and visually rich content ( Doloi 2024 ) . Microlearning is often described as a flexible approach that enables learners to target relevant tasks efficiently ( Kohnke et al. 2024 ) . Consequently, researchers call for studying microlearning through platforms like TikTok and YouTube Shorts to expand access to educational content ( Aslan 2024 ) . However, most existing work focuses on human-curated content, leaving limited empirical evidence on the pedagogical value and learner experience of AI-generated short-form videos. Generative AI in Education. Generative AI has created new ways for generating and personalizing educational content ( Baidoo-Anu and Ansah 2023 ; Niu and Xue 2023 ; Qadir 2023 ; McDonald et al. 2025 ; Liarokapis et al. 2024a ; Liarokapis et al. 2024b ; Constantinides et al. 2024b ) . Tools such as ChatGPT have prompted experiments with AI-driven teaching assistants, automated feedback, and content generation. Boumalek et al. ( Boumalek et al. 2024 ) outlined how generative models can produce concise educational content through stepwise processes, boosting engagement and scalability. Reviews highlighted benefits such as improved accessibility, personalized support, and learning outcomes ( Bettayeb et al. 2024 ; Mittal et al.