When researchers at ETH Zurich asked an image generator for visuals to explain “operations management,” it returned photos of surgery. The word “operations” sent it straight to the hospital.
That small glitch sums up where AI sits in educational video right now. It’s fast, cheap, and often good enough, until it confidently teaches the wrong thing. Schools, universities and training teams are now weighing educational video production Melbourne studios against do-it-yourself AI tools. For them, the useful question is narrower than “AI or human?” It’s which parts of a lesson can safely be generated, and which still need a crew, a subject expert and a careful edit.
What a 447-Person Experiment Found About AI Teaching Videos
The ETH Zurich team, led by Torbjørn Netland, rebuilt four short management lectures with generative AI. ChatGPT wrote the scripts, Midjourney and DALL-E made the visuals, and Movio (now HeyGen) turned a single photo of each lecturer into a talking avatar. The team then showed 447 online participants a mix of AI and human-made versions and tested them afterwards. The study appeared in Computers & Education in 2025.
The results are more interesting than a simple win or loss:
- Exam scores were statistically the same for both versions.
- Viewers rated their learning experience slightly higher for the human-made videos, about 4% on average.
- Each AI video took 3 to 4 hours to make, and none took more than 8. Each human version needed around 3 hours of scripting, an hour of filming and 12 to 24 hours of production.
Look closer and the picture splits. For two of the topics, viewers mildly preferred the AI version. For the other two, a problem-solving method and an inventory formula, they disliked it more strongly. The authors also flag their own limits: the videos ran about three minutes, covered simple and well-settled concepts, and were watched by paid online participants instead of enrolled students.
So AI can carry a short, clear-cut concept. Nobody has shown yet that it can carry a whole unit, a contested topic or a hands-on skill.
The Shortcut Victorian Government Schools Can’t Take
Here’s the catch for schools in this state. The cheapest trick in that experiment was cloning a real teacher into an avatar from one photo. That runs straight into Victoria’s Department of Education policy on generative AI.
The policy tells schools to direct staff and students not to upload photos, audio or video of students, staff or parents into generative AI tools. It also tells them not to create media in those people’s likeness. Staff can’t use generative AI to communicate with students and parents in ways that replace a teacher’s own voice and professional judgement. Any tool that needs personal details beyond a school email requires opt-in consent from parents and carers. Schools should also disclose AI use when it affects others.
In practice, that rules out a “digital twin” of your Year 9 science teacher. Stock avatars and AI narration sit in greyer territory, and anything that collects student data needs consent paperwork first. Independent schools, Catholic schools and universities work under their own frameworks, so check yours before signing up for a subscription.
There’s a quieter risk too. Staff sometimes paste lesson notes, student work samples or assessment data into free video tools. That’s shadow AI inside the school, and the department’s privacy rules treat that kind of upload as a potential breach.
Where AI Actually Helps a Video Project
None of this makes AI useless. It changes where AI sits in the pipeline. The smartest production teams use it for the grunt work and keep people on anything a learner might believe.
| Production task | Where AI helps | What a human must own |
| Script | Fast first drafts, simpler wording for younger viewers | Accuracy, curriculum fit, tone |
| Storyboard | Rough frames to test ideas quickly | Picking visuals that actually teach |
| Narration | Temporary voice for timing checks | Final voice, consent and rights |
| Captions | Auto-generated first pass | Fixing technical terms and names |
| Versions | Translations, shorter cuts, different audiences | Checking each version still says the right thing |
Captions deserve a second look. Auto-captioning can stumble on technical vocabulary and local place names, and an error in a caption becomes an error in someone’s notes. Victoria’s AI policy already reminds schools of their duties under the Disability Discrimination Act and the Disability Standards for Education when adopting AI tools. Sloppy captions undercut exactly those duties.
Narration raises its own questions. If a studio offers an AI voice, ask whose voice it was trained on and what that person agreed to. AI voice-over contract traps catch narrators and clients alike.
Formats That Still Need a Camera and a Crew
Some educational videos depend on things AI can’t fake honestly.
Campus tours and open days. A prospective student wants to see the real library, the real labs and the real walk from the tram stop. A generated campus is a misleading one.
Procedural and clinical skills. Hand positions, equipment handling and safety steps need real footage of a real expert. A plausible-looking AI demonstration that gets one step wrong is worse than no video at all.
Staff training on live systems. Recording your actual software, site or workflow beats any generic stand-in. Melbourne studio Burninghouse says it built a series of IT training videos for ANZ that reached more than 4,000 bankers across Asia Pacific. At that scale, one inaccurate step gets repeated thousands of times.
Case studies and testimonials. These are real people describing real outcomes, and synthesising them crosses an honesty line quickly.
Animation sits in the middle. When the subject is invisible or abstract, like how a virus spreads or how money moves through a super fund, motion graphics can show what a camera can’t. Burninghouse lists an animated training piece for the Victorian Government among its work. In projects like these, AI can speed up early drafts while illustrators and instructional designers lock down the final frames.
How to Choose an Educational Video Partner in 2026
Years of experience and a polished reel still matter. Now add a few AI questions to the briefing.
| Question to ask | Why it matters |
| Where do you use generative AI, and will you disclose it? | You may need to tell learners or parents |
| Who checks every fact before delivery? | AI drafts sound confident even when they’re wrong |
| How do you handle likeness, voice and filming consent? | Victorian schools have firm rules on images of students and staff |
| Are captions checked by a person? | Accessibility duties and learner trust both depend on it |
| What’s in the quote: scripting, revisions, animation, versions? | Clear scope prevents surprise costs |
On budget, cost follows complexity. A short explainer built from existing graphics costs far less than a multi-module series with scripting, animation, location filming and several feedback rounds. If a studio uses AI to cut drafting time, it’s fair to ask whether that saving shows up in the quote.
The Production Process, With AI Checkpoints
Most projects still move through the same stages. What’s changed is where the checks sit.
- Brief and learning goal. Agree on the audience, the one thing viewers should be able to do afterwards, and any AI rules your institution follows.
- Script and storyboard. AI can rough these out, but a subject expert signs off every claim.
- Filming or animation. Real people and places get filmed with consent forms in hand. Abstract processes get animated.
- Edit and captions. Captions get checked line by line, especially the terminology.
- Delivery and versions. Cut-downs, translations and LMS-ready files are each rechecked before they go out.
A realistic schedule leaves room for step 2. Rushing the script is where both humans and AI slip up.
So Where Does That Leave Melbourne Educators?
The ETH Zurich result is hard to ignore. For a short, simple concept, an AI-built video taught about as well as a human one, in a fraction of the time. Ask harder questions and the answer gets murkier. Would it hold up for a 20-minute unit on a contested topic? For a skill students practise with their hands? For a school that can’t put a teacher’s face into the tool at all?
Maybe the better split is to let AI make the drafts and let people make the moments learners remember. Where your institution draws that line is still open, and it’s worth settling before the next video brief goes out.
FAQs
Q.1 Can AI-generated teaching videos replace professional production?
For short, well-defined concepts, ETH Zurich researchers found AI videos produced similar test scores, though viewers slightly preferred the human-made ones. Nobody has shown the same for longer, complex or hands-on content.
Q.2 Can Victorian government schools use AI avatars of teachers?
The Department of Education’s generative AI policy directs staff not to upload or generate media in the likeness of students, staff or parents, which effectively rules out teacher avatars. Independent schools and universities should check their own sector’s rules.
Q.3 What drives the cost of educational video production in Melbourne?
Length, format (live action or animation), scripting effort, the number of revision rounds and how many versions you need all drive cost. Ask for each stage to be priced separately.
Related: I Tested an Uncensored AI Video Extender—Here’s What Happened
