Students Prefer Personalized, AI-Generated Educational Videos over Non-Personalized, Human-Recorded Videos

Synthesis: Tomlinson, Black, Patterson, van der Hoek, Ferguson, and Bietz (2026) field-deploy personalized AI-generated educational videos as the primary instructional modality in a large online undergraduate course and ask students to rank four video types defined by two crossed dimensions — personalization (personalized vs. non-personalized) and source (human-recorded vs. AI-generated). Across two offerings (493 respondents), personalization outweighed human presence: students preferred AI-generated personalized videos over non-personalized human-recorded videos (mean rank 2.26 vs. 2.69, p < .001), and 88.4% ranked some personalized video first versus 73.8% for human-recorded. Human-recorded personalized videos were ranked highest overall, but they were a hypothetical condition students never experienced. The authors read the results as a turning point: personalized relevance and conciseness now compensate for, and sometimes surpass, the missing human presenter — pointing toward a complementary model where human instructors provide expertise and social connection while generative AI supplies scalable personalization.

Key Findings

1. Personalization is a stronger driver of preference than human presence. In a direct comparison, AI-generated personalized videos beat non-personalized human-recorded videos (mean rank 2.26 vs. 2.69; Wilcoxon signed-rank, p < .001) — stable across both cohorts. Grouped by personalization, 88.4% of students ranked a personalized video (human or AI) as their top choice (binomial p < .001); grouped by source, only 73.8% ranked a human-recorded video first. The magnitude of the personalization effect substantially exceeded the effect of human presence.
2. Full preference ordering. Averaging across both offerings, the four types ranked: hypothetical human-recorded personalized (mean rank ~1.5) > AI-generated personalized (~2.26) > human-recorded non-personalized (~2.69) > AI-generated non-personalized (~3.54). Human-recorded personalized videos were highest, but because none existed, this condition mixes "informed imagination" with lived experience and carries an unknown imagination bias.
3. AI video design for personalization. The team built a custom pipeline with human oversight at multiple stages: the instructor authored an LLM "ethos" prompt; AI selected 100 topics; for each topic three video versions were generated tailored to business, technology, or society/biology majors (chosen from an AI analysis of enrolled students); each shared three core-content paragraphs and diverged in the last two for domain-specific examples. Scripts were rendered by a HeyGen avatar cloned from the instructor (straight/left/right facing), composited with images, bullet points, titles, and music into 3–5 minute videos, reviewed for correctness, then pushed to YouTube and into Canvas pages.
4. Student-stated benefits and drawbacks. Open-ended responses highlighted relevance (videos "catered to me based off of my interests"), clarity, consistency, and conciseness (more direct than human lectures "filled with jargon/ramble") as benefits. Drawbacks centered on naturalness and expressiveness — the limits of the AI avatar — plus, for a subset of students, principled ethical objections to AI use (environmental impact, labor, automation, academic integrity) that technical improvement alone cannot address.
5. Caveats. Limitations include the experiential asymmetry across conditions (three were experienced/largely hypothetical/parallel), potential imagination bias and contrast effect (many students' baseline was generic human lecture video, which may inflate the personalization preference via novelty), demand characteristics, a single general-education online course at one university dominated by seniors, and a sustainability-and-computing subject that may bias toward technology-oriented students. Objective learning-outcome data are reported as forthcoming.

Connected Concepts

• Personalized Learning — the central construct; personalized relevance drove preference
• Video Education — Video in Education: AI-generated, personalized, and analytics of video learning
• Generative AI — the LLM (GPT-4o, Claude) and HeyGen avatar video-generation pipeline
• Online Teaching And Learning — video-based instruction as the primary modality in online courses
• Student Engagement — personalization as a driver of engagement and preference
• Adaptive Learning — scalable personalization of pacing/content at course scale
• Higher Ed — the large undergraduate general-education course context
• Pedagogical Agent — the AI avatar as a virtual pedagogical agent/presenter
• Multimodal — AI-generated video combining avatar, image, bullet, and audio modalities

Connected Articles

• AI Generated Instructional Videos Computing Ed — student perceptions and preferences of AI-generated instructional videos in computing education
• AI Video Dual Gatekeeping 2026 — dual gatekeeping for pedagogically grounded AI content (video) creation
• Courseblueprint Adaptive Video Generation — CourseBlueprint: adaptive pedagogical video generation grounded in course corpora
• Face Value How Avatar Identity Shapes Epistemic Trust In AI Mediated Learning — how avatar identity shapes epistemic trust in AI-mediated learning
• Engagement Assessment Video — engagement assessment in video learning
• Savvy Student Attention Video Learning — student attention visualization for video-based learning
• AI Psychotherapy Training Avatars — AI avatar use in (psychotherapy) training contexts

Citation

Tomlinson, B., Black, R. W., Patterson, D. J., van der Hoek, A., Ferguson, J., & Bietz, M. J. (2026). Students prefer personalized, AI-generated educational videos over non-personalized, human-recorded videos. Scientific Reports, 16, 21804.