Bill Tomlinson, Rebecca Black, André van der Hoek, Julie Ferguson, Matthew Bietz, and I wrote a paper that was accepted to Springer Nature Journal, Scientific Reports, in which we ask how students weigh the benefits of personalization against the value of a human instructor in educational video. I led the AI video production work for this one.
What did we find?
- Across two offerings of a large online undergraduate course (493 respondents), students preferred AI-generated personalized videos over human-recorded non-personalized videos by a substantial margin (mean rank 2.26 vs. 2.69; Wilcoxon signed-rank test, p < 0.001).
- 88.4% of students ranked a personalized video (either AI-generated or human-recorded) as their top choice, indicating a strong and robust preference for personalization.
- When source was held constant, 73.8% of students preferred human-recorded over AI-generated content — so human presence still matters, but the magnitude of the personalization effect substantially exceeded the effect of human presence.
- In open-ended responses, students highlighted the relevance and conciseness of personalized AI videos, alongside concerns about naturalness, expressiveness, and uncanny valley effects.
Personalization is a well-established driver of student engagement, yet delivering individualized instruction at scale remains a challenge in online education. Recent advances in generative AI make scalable personalization feasible, but AI-generated educational videos are often perceived as inferior to human-recorded content. This tension raises the question: how does the value of personalization compare to that of human presence?
We investigated this question through a field deployment in two offerings of a large undergraduate online course. AI-generated personalized videos served as the primary instructional modality, alongside a smaller set of non-personalized human-recorded and AI-generated videos. At the end of the course, students ranked preferences across personalized and non-personalized formats and human-recorded versus AI-generated content.
For each of 100 topics in the course, our pipeline generated three variants of a five-paragraph video, one tailored to business, one to technology, and one to society/biology, using an LLM-generated script anchored to the instructor’s “ethos” prompt, a HeyGen avatar trained on a recording of the instructor, and dynamically composited images, bullet points, titles, and credits. Students chose which variant they wanted to watch on Canvas. The 300 videos received over 17,500 views across the two quarters.
Rather than framing the future of education as a choice between AI and instructors, these findings point toward a complementary model: human instructors provide expertise, mentorship, and social connection, while AI systems extend their reach by enabling personalized instruction that would otherwise be infeasible at scale. There are real caveats — the human-recorded personalized condition was hypothetical, novelty effects may be inflating the personalization signal, and preference is not the same thing as learning outcome (a companion study on that is in the works). But the result is suggestive: relevance and contextual fit can compensate for, and in some cases surpass, the absence of a human presenter.
Demo video:
First Generation Course Content Video:
Related papers by others (also see references in the paper): Comparing human-made and AI-generated teaching videos: An experimental study on learning effects, AI Tutoring Outperforms In-Class Active Learning: An RCT Introducing a Novel Research-Based Design in an Authentic Educational Setting, University students describe how they adopt AI for writing and research in a general education course
(permanent, open-access, local copy)
C.V.: JR-20