Quick Answer
A peer-reviewed 2026 study at two US universities had 170 computing students watch AI-generated instructional videos made with Knowlify. Students rated the videos 4.52 out of 5 for production quality and scored 86% on a knowledge test, half could not tell the videos were AI-generated, and 68% still preferred videos recorded by their instructor.
TL;DR: Researchers at the University of Florida and Worcester Polytechnic Institute asked 170 computing students to watch three short AI-generated instructional videos built with Knowlify, then surveyed what they thought. Students rated the videos 4.52 out of 5 for production quality, 4.43 for accuracy, and scored an average of 86% on a knowledge test afterward. Half of them could not tell the videos were AI-generated at all. But when asked about AI video replacing their instructors, support collapsed: 68% said they would not prefer AI video over instructor-recorded video, and 62% said they would not trust it equally. The takeaway for anyone building a video program is narrow and useful. AI video is ready for supplemental, foundational, and visual content. It is not ready to be the lecture.
See also: learning science principles
The Study in One Table
The paper, Student Perceptions and Preferences Regarding AI-Generated Instructional Videos in Computing Education, is a descriptive post-test survey study published in 2026 by researchers at the University of Florida and Worcester Polytechnic Institute.
| Study element | Detail |
|---|---|
| Students analyzed | 170 (143 at a large public university, 27 at a polytechnic university) |
| Recruitment | Upper-level Data Structures and Algorithms and Software Engineering courses |
| Response rate | 45% (215 of 475 enrolled, 170 retained after exclusions) |
| Video topic | Markdown markup language for GitHub README files |
| Videos watched | Three videos, 9 minutes 31 seconds total |
| Video format | Animated visual walkthroughs, deliberately no avatars or talking heads |
| Tool used | Knowlify, generating from GitHub's official Markdown documentation as a PDF |
| Survey | 24 questions on Qualtrics, 19 analyzed, 15.5 minutes average completion |
| Prior familiarity | 64% and 59% of students at each site were not at all or only slightly familiar with Markdown |
One design decision matters more than the rest. The researchers deliberately excluded AI avatars and talking heads to remove social presence as a confounding variable. Earlier research had found that students perceived AI-generated faces and voices as distracting and uncanny, which made it impossible to separate a reaction to the format from a reaction to the synthetic presenter. By using animated explainers instead, this study isolated the question of whether AI-generated instruction works, rather than whether AI-generated humans are unsettling.
Students were also not told in advance that the videos were AI-generated, so their quality ratings were not biased by the label.
Finding 1: Students Rated the Videos Highly on Every Quality Measure
Across all 170 students, ratings were strongly positive on a five-point agreement scale.
| What students were asked | Mean rating | Agreement |
|---|---|---|
| The videos were professionally produced | 4.52 | 92% |
| The videos helped me understand Markdown | 4.39 | 92% |
| The videos explained Markdown accurately | 4.43 | 88% |
| The videos were engaging | 4.20 | 80% |
| I would use similar videos for future learning | 4.06 | 76% |
Production quality scored highest at 4.52 out of 5. This is a meaningful result because production quality is exactly where earlier AI video research found the most friction. In prior studies, university instructors reported audio-visual mismatches and inaccurate content in AI-generated course videos, and students described AI presenters as uncanny. Neither showed up here.
Engagement at 4.20 and willingness to reuse at 4.06 were the two lowest scores in the set, which tracks with the open-ended comments. Some students described the narration as sounding samey across the three videos, and a few said they tuned out over time.
Finding 2: Half the Students Could Not Tell the Videos Were AI-Generated
This was the only item in the entire survey without consensus, and it produced the highest variance of any question.
- 50% agreed they could clearly tell the videos were AI-generated
- 35% disagreed
- 15% were neutral
- Mean rating of 3.31 out of 5
Put plainly, after watching nearly ten minutes of instructional content, roughly half of a room full of upper-level computer science students could not reliably identify AI-generated video. These are not casual viewers. They are technically sophisticated users in advanced CS coursework, primed to notice artifacts.
The researchers attribute this to two factors: the general advance in AI video generation quality, and the absence of an avatar. Synthetic faces and voices are where detection usually happens. Remove the presenter, and an animated explainer built by AI looks like an animated explainer.
Finding 3: Students Actually Learned the Material
Perception scores are worth little if nothing is retained, so the survey included a five-question knowledge test covering both factual and interpretive items from the video content.
Students at both institutions averaged 4.3 out of 5, or roughly 86%. Median score was 5 out of 5 at the larger institution. This held despite most students reporting little or no prior familiarity with Markdown before watching.
Viewing logs backed up the self-reports rather than relying on them. Around 90% of students said they watched all three videos, and platform logs showed average watch coverage of 76% to 81% per video across both sites. Students were genuinely watching, and they genuinely learned.
The researchers are careful here, and so are we: this measures short-term knowledge acquisition, not long-term retention or transfer to new problems.
Finding 4: Support Collapses When AI Video Replaces the Instructor
Here is the part that should shape how any organization deploys this technology. The same students who rated the videos 4.52 for quality were far more guarded about AI video in their coursework generally.
| Statement | Agree | Disagree | Mean |
|---|---|---|---|
| I am comfortable with AI videos as instructional material in my courses | 46% | 39% | 3.00 |
| I would prefer AI videos over videos recorded by my instructor | 10% | 68% | 2.06 |
| I would trust information in an AI video as much as a human-presented one | 17% | 62% | 2.24 |
Comfort with AI video as course material split almost evenly. But preference and trust both drew clear opposition. Only one student in ten preferred AI video to instructor-recorded video, and fewer than one in five trusted it equally.
Read those two blocks of findings together and the picture is consistent rather than contradictory. Students are not rejecting the format. They are rejecting the substitution. AI video that supplements instruction is welcome. AI video that displaces an instructor is not.
Where Students Said AI Video Belongs
The researchers coded 163 open-ended responses into themes. These are the contexts students identified as appropriate, with the share of students who named each one.
| Use case | Students | Share |
|---|---|---|
| Simple or foundational topics such as vocabulary and basic syntax | 44 | 27% |
| Topics that benefit from visualization such as data structures and processes | 24 | 15% |
| Topic summaries, overviews, and exam review | 20 | 12% |
| Supplemental material that reinforces a concept a second way | 20 | 12% |
| Replacing dense written documentation | 11 | 7% |
| Tailored and multilingual learning for students who need to catch up | 8 | 6% |
| Improving presentation clarity, including narration over instructor slides | 4 | 2% |
| Serving settings that lack funded instructors | 6 | 4% |
| Reducing instructor workload on routine content | 6 | 4% |
Two of these deserve attention from anyone running a content operation.
The visualization case was specific and enthusiastic. One student wrote that AI videos "work extremely well for explaining data structures like Linked Lists and Trees since they can add elements with an animation, show what traversal would look like." That is a description of animated explainer video doing something static text cannot.
The documentation replacement case is the one with the clearest commercial parallel. As one student put it, "watching a short video on something is a lot less daunting than reading a tedious article." That is the same argument for turning technical documentation into video, and this study is the first empirical evidence we have seen that the target audience agrees.
Where Students Said It Does Not Belong
| Poor fit | Students | Share |
|---|---|---|
| Complex, advanced, or nuanced topics | 50 | 31% |
| Subjective topics such as philosophy, ethics, art, and public speaking | 22 | 14% |
| Replacing instructor-led lectures entirely | 26 | 16% |
| Interactive learning such as debugging, algorithm design, and seminars | 14 | 9% |
| Niche topics with limited source material online | 6 | 4% |
The boundary students drew is sharp. Complex material needs elaboration, back-and-forth, and someone who can respond to the specific shape of a student's confusion. As one student explained, discussions "require dynamic back-and-forth perspective sharing and real-time responses to student viewpoints." Nobody claimed a video could do that.
What Students Are Worried About
Concerns were coded from the same 163 responses. Nine students (6%) reported no concerns at all.
| Concern | Students | Share |
|---|---|---|
| Inaccurate or hallucinated information | 94 | 58% |
| Lack of depth, oversimplification, skipped steps | 22 | 13% |
| Lack of engagement, monotone delivery, pronunciation errors | 21 | 13% |
| Fewer interactions with instructors and TAs | 16 | 10% |
| Undermines the value of tuition | 17 | 10% |
| Missing instructor insight and personal experience | 9 | 6% |
| Broader societal impacts including instructor pay and energy use | 9 | 6% |
| No thorough fact-checking by an educator | 7 | 4% |
| Cannot ask follow-up questions | 6 | 4% |
| Instructors may reduce effort and offload teaching | 5 | 3% |
Accuracy dominates everything else. At 58%, it was named more than four times as often as the next concern. One student summarized the asymmetry neatly: "although professors can make mistakes, my experience of AI is that it will."
There is a subtler concern in that table worth pulling out. Seven students noted that high production quality can itself be a risk, because a polished video can "lull a false sense of security" and make errors easier to miss. Given that this study also found students rated production quality at 4.52 and half could not detect AI generation, that warning is well placed. The better the output looks, the more the human review step matters.
What This Means If You Are Building Video
The study was run in a computing classroom, but the findings transfer cleanly to corporate training, customer education, and documentation.
Deploy AI video where students already said it works. Foundational concepts, syntax and procedure, visual processes, summaries and refreshers, and supplemental explanations of material that also exists in writing. This maps almost exactly onto the highest-volume, lowest-differentiation content in most training video libraries and most new-hire onboarding programs.
Keep a human in the loop and say so. Accuracy was the top concern by a wide margin, and four percent of students specifically worried that nobody had reviewed the video end to end. In the study itself, each video included roughly 10 minutes of human verification time. Build that step into your process and tell users it happened.
Do not position AI video as a replacement for a person. The data on this is unambiguous. Comfort sits at 46% when AI video is described as instructional material and drops to 10% when it is framed as a substitute for the instructor. The framing changes the reception more than the content does.
Expect the format to carry more weight than you think. The single most consequential design choice in this study was removing the avatar. Animated explainers scored 4.52 on production quality in a literature where avatar-based videos had previously been called distracting and uncanny. If your audience is skeptical, animation is the safer format.
How the Videos Were Made
For teams evaluating the production economics, the paper documents the workflow precisely.
The researchers uploaded GitHub's official Markdown formatting documentation as a PDF to Knowlify and used a one-shot prompt to generate each video. The platform produced a narration script, a scene-level storyboard, and voiceovers, all reviewable and editable before final render.
| Production metric | Result |
|---|---|
| Planning, reviewing, and configuring | 8 to 11 minutes per video |
| System generation time | 3 minutes 48 seconds average |
| Human verification | About 10 minutes per video |
| Total time per 3-minute video | At most 25 minutes |
| Cost per video at study-period pricing | Roughly $50 to $60 |
Twenty-five minutes of total human and machine time per finished three-minute video is the number worth holding onto. Traditional instructional video production is typically quoted per finished minute in the hundreds to thousands of dollars and measured in days or weeks. For a fuller cost comparison framework, see our guide on measuring the ROI of AI video.
Limitations Worth Stating
The authors are explicit about the boundaries of this work, and the findings are more useful when those boundaries are respected.
It was a descriptive study, not a controlled comparison. It did not test whether AI videos outperform human-recorded videos or written documentation. The 86% quiz average shows learning happened, not that it happened better.
The students were enrolled in upper-level CS courses. Even though most had little Markdown experience, they had substantial background in programming and technical notation, which likely made the material easier to absorb. Results may differ for introductory learners.
The content was simple and short. Three-minute videos on basic syntax are close to the ideal case for this format, and the students themselves flagged complex material as a poor fit. Nothing here says AI video handles advanced topics well.
The qualitative themes come from one coding process and are open to other interpretations, which the authors acknowledge and address by publishing their codebook process and representative quotes.
Disclosure
Knowlify provided the research team with early access to the platform and complimentary video credits for this study. As stated in the paper's acknowledgements, Knowlify had no role in the study design, data collection, analysis, interpretation, or the decision to publish. We are writing about this research because it is independent, and we think the parts that are unflattering to AI video are as interesting as the parts that are not.
Key Takeaways
- 170 computing students at two US universities watched three AI-generated instructional videos built with Knowlify and rated them 4.52 out of 5 for production quality, 4.43 for accuracy, and 4.39 for helpfulness.
- Half the students could not clearly tell the videos were AI-generated, the highest-variance item in the survey.
- Students averaged 86% on a post-viewing knowledge test despite most having little or no prior familiarity with the topic.
- Support drops sharply at substitution: 68% would not prefer AI video over instructor-recorded video and 62% would not trust it equally.
- Students endorsed AI video for simple topics (27%), visual topics (15%), summaries (12%), and supplemental material (12%).
- Accuracy is the dominant concern at 58%, named more than four times as often as the next issue.
- Removing AI avatars appears to be a significant reason the videos were received well, since prior research found synthetic presenters distracting and uncanny.
- Each three-minute video took at most 25 minutes of combined design, generation, and human verification time.
FAQ
Do students like AI-generated instructional videos?
Yes, when the videos supplement rather than replace instruction. In a 2026 study of 170 computing students, 92% agreed the AI-generated videos were professionally produced, 92% said the videos helped them understand the material, and 88% said the content was accurate. However, only 10% said they would prefer AI video over videos recorded by their own instructor, so approval is conditional on the videos being an addition rather than a substitution.
Can students tell if a video is AI-generated?
Often not. In the 2026 study, only 50% of upper-level computer science students agreed they could clearly tell the videos were AI-generated, while 35% disagreed and 15% were neutral. This was the only item in the survey without consensus. The videos used animated visual walkthroughs rather than AI avatars, which the researchers suggest is a key reason detection was so difficult, since synthetic faces and voices are the usual giveaway.
Do students learn as well from AI-generated videos?
Students in this study learned the material well, averaging 4.3 out of 5 (roughly 86%) on a knowledge test after watching, despite most reporting little or no prior familiarity with the topic. Platform logs confirmed they watched 76% to 81% of each video on average. Important caveat: the study was descriptive rather than comparative, so it measured short-term knowledge acquisition and did not test AI video against human-recorded video or written documentation.
What topics work best for AI-generated educational videos?
Students identified four categories. Simple or foundational topics such as vocabulary and basic syntax were named by 27% of students. Topics that benefit from visualization, such as data structures and step-by-step processes, were named by 15%. Topic summaries and exam review were named by 12%, and supplemental explanations that reinforce a concept a second way were also named by 12%. Students specifically flagged complex, subjective, and interactive topics as poor fits.
What are students most concerned about with AI videos in education?
Accuracy, by a wide margin. 58% of students raised concerns about inaccurate or hallucinated information, more than four times the rate of the next concern. After that came lack of depth (13%), lack of engagement (13%), reduced interaction with instructors (10%), and the sense that AI video undermines the value of tuition (10%). Only 6% of students reported no concerns at all.
Should AI videos replace lectures?
The students in this study said no, clearly. 68% disagreed that they would prefer AI videos over instructor-recorded videos, 62% disagreed that they would trust AI video equally, and 16% specifically named replacing instructor-led lectures as an inappropriate use. The researchers conclude that AI videos can play a meaningful role in education without displacing the instructor presence that students value.
How long does it take to make an AI instructional video?
In this study, each three-minute video took at most 25 minutes end to end. That breaks down as 8 to 11 minutes of planning, reviewing, and configuring, 3 minutes 48 seconds of average system generation time, and about 10 minutes of human verification. The researchers generated the videos by uploading GitHub's official Markdown documentation as a PDF to Knowlify and using a single prompt.
Why did the researchers avoid AI avatars in the videos?
To remove social presence as a confounding variable. Prior research had found students perceived AI-generated appearances and voices as distracting and uncanny, even when the script was identical to a human-produced version, which made it impossible to separate reactions to AI instruction from reactions to synthetic humans. By using animated visual walkthroughs with no talking heads, this study isolated the question of whether AI-generated instruction works. The high production-quality ratings suggest the choice mattered.