On this page
AI video generators can now turn a typed script into a lecture delivered by a realistic avatar with a synthetic voice. The early research on whether they work is small but fairly consistent, and it points to one thing worth checking: the explanation, not the face.
Can AI-generated video lecturers teach as well as humans?
The largest of the studies here is by Torbjørn Netland and colleagues at ETH Zurich, published in Computers & Education in 2025. They made human-made and AI-generated versions of four short teaching videos on production management concepts. In an online experiment, 447 participants were split into two groups; each watched two human-made and two AI-generated videos, with the versions swapped between groups. There was no statistically significant difference in exam results between the AI-generated and human-made videos, as Netland describes the study. Participants did enjoy the human-made videos more.
A smaller study pointed the same way. Selay Arkün-Kocadere and Şeyma Çağlar Özhan compared an AI-generated instructor with a human instructor in a study involving 108 undergraduates, reported in the International Review of Research in Open and Distributed Learning in 2024. Video engagement was higher with the human instructor, but instructor type had no significant effect on academic performance. In a focus group, students said the AI-generated instructor caused distraction, discomfort and disconnectedness, feelings they could set aside when the topic was interesting or when they watched with the intention of learning.
Does a synthetic voice change how much students trust a video?
It can change trust without changing learning. In a 2020 study in Computers & Education, Erin Chiou, Noah Schroeder and Scotty Craig found that the quality of a virtual human's voice did not significantly influence learning, but it did influence trust and learners' other perceptions of the virtual human.
In a 2022 study in the Journal of Computer Assisted Learning, Zhongling Pi and colleagues compared a human instructor with virtual instructors that combined human or AI voices with AI-generated images. The authors reported that a virtual instructor with an AI voice and a likable AI-generated image had a positive effect on learning performance and motivation, without reducing attention to the lesson materials.
So far, learning holds up while comfort and trust can drop, and a student who tunes out may stop watching before any test.
What do these studies not tell you yet?
Length: the IRRODL groups watched two short videos over two weeks, and the Netland study used four short videos. A full syllabus is a different test.
Learners: undergraduates and adult online participants, not school students preparing for JEE Main or board exams.
Subjects: management and undergraduate course content, not long physics or maths derivations, where one wrong sign ruins the answer.
Time: short experiments cannot tell you what you will still remember months later.
How do you check an AI video lecture before you rely on it?
Use this five-point check on the first video from any new channel or app.
Who wrote the script? Look for a named teacher, institution or textbook. The face tells you nothing; the script teaches.
Is it labelled as AI? A realistic avatar presented as a real teacher is a reason to be careful.
Check one step by hand. Pause at a worked example and redo one line. If a sign, unit or formula is wrong, drop that video for that topic.
Listen for notation read aloud wrongly. For example, "sin² θ" read as "sin 2 theta", or a Greek letter skipped. If the voice and the board disagree, check the textbook before trusting either.
Pause and predict. Before each new step, pause and say what comes next. If you cannot, the video is washing over you, which is the kind of disengagement the studies measured.
Here is step 3 on a real case. Say an AI video on projectile motion gives the range as R = u² sin 2θ / g and says the range is greatest at 30°. Check it: sin 2θ is largest, equal to 1, when 2θ = 90°, so the greatest range is at θ = 45°, where R = u²/g. One line of checking catches the error.
Where does an AI tutor fit in?
A video, human or synthetic, cannot notice that you are lost at step three. That is the gap Fermi's AI tutor is built for: it holds a line of reasoning across a session, notices where a student is stuck, and decides whether to explain, ask or show. For ideas that are easier to see than to read, such as how the launch angle changes the range, the Fermi Simulation Lab has interactive simulations.
Next time an avatar teaches you something, pause at one step and check it yourself. That habit matters more than who, or what, is on screen.
