Do Flaws in AI Avatars Hurt Learning?
Visual flaws in AI avatars may matter when they obscure an explanation, misrepresent a task, or undermine trust. But a designer noticing a flaw does not establish that it harms learning. The research discussed here offers reasons to reconsider automatic rejection of synthetic presenters, while leaving questions about specific defects unresolved. Evaluation should distinguish visual discomfort from effects on understanding, performance, and trust.
I’ve spent years training myself to spot what’s wrong with a training visual before a single learner ever opens the course. That’s the job. Lately, what I keep spotting is AI. A hand with one knuckle too many. A smile that lands half a beat after the sentence that earned it.
I notice these things instantly. What I hadn’t established was whether the flaws I noticed were actually getting in the way of learning. I’d been letting my reaction answer a question I hadn’t really tested.
Long before instructional design, I worked as a technical writer. The hardest skill I built wasn’t writing clearly. It was learning to set my own view aside and try to see through the reader’s eyes. That person didn’t know what I knew or necessarily notice what I noticed. I’m still not great at that.
That leaves me with a different question about AI avatars in training. How do I use my trained eye without mistaking it for evidence of the learner’s experience? A flaw can deserve attention without being a reason to reject the whole approach. The decision needs to consider what it does to understanding, performance, and trust.
Chris Anderson makes a related point in his talk on public speaking. He describes building an explanation from concepts the audience already understands, rather than assuming they share the speaker’s knowledge. That’s advice about explaining ideas, not evaluating avatars, but I recognize the discipline: start where the audience is. [1]
So I went looking beyond my own reaction. The research I found helped, although it didn’t answer quite the question I’d started with.
01
RESEARCH
What the Research Says About Animated Instructors
In a 2024 study published in PNAS, Chanyuan Gu and colleagues compared video lectures with human instructors, an animated instructor, and no visible instructor. They used eye-tracking and functional MRI, then assessed comprehension, recall, and transfer. [2]
In the human-instructor comparison, learners performed better with the presenter visible than with narrated slides alone. But the comparison that interested me most involved 18 learners watching a smiling human instructor and 16 watching a smiling animated one. The researchers found no statistically significant differences between those groups on the three learning measures. [2]
That gives me a reason to question my automatic preference for a human presenter. It doesn’t establish that the two options are equivalent. These were small groups, and failing to detect a difference isn’t the same as establishing that there isn’t one.
The result also doesn’t tell me whether learners noticed an odd hand or an awkward expression. Those weren’t the questions being tested. It does make me less comfortable assuming that a synthetic presenter must be a worse one.
02
LIMITS
Where My Confidence Runs Out
The animated instructor inwas stylized, and all versions used recordings from the same human speaker. This wasn’t a test of a photorealistic AI presenter delivering synthetic speech, or a comparison between otherwise identical videos with and without visual defects. [2]
Those differences matter to the decision I’m trying to make. I can’t take a finding about one presentation and apply it to every avatar that resembles it.
Masahiro Mori’s uncanny valley offers a way to think about one possible problem. Proposed in 1970, it describes a dip in our affinity for something as it approaches a human appearance without quite achieving it. An almost-human figure may feel more unsettling than an obviously artificial one. [3]
That helps me name the discomfort. It doesn’t tell me how strongly a particular learner will feel it, or whether it will interfere with learning. I shouldn’t dismiss that possibility any more than I should assume it.
There is also a difference between a cosmetic defect and an instructional error. An odd-looking hand resting on a desk raises one question. A hand demonstrating the wrong way to perform a task raises another. I don’t need a learner trial to justify correcting an inaccurate demonstration. The content still has to be right.
03
DISCLOSURE
What About Telling People the Instructor is AI?
The disclosure study I found doesn’t settle what telling learners will do to trust or learning. It does give me another example of why I should test expectations rather than treat them as facts.
Boele Visser and colleagues at Leiden University tested disclosure in a Roblox game. Their preprint reports an analysis of 72,903 players. Telling participants that their companion was AI reduced average time per level by about 1.5 seconds. The researchers interpreted that statistically significant but small difference as an increase in play intensity. [4]
The companion was a simply programmed game character. The experiment didn’t directly measure trust, comprehension, or retention. I can’t use it to say that disclosure builds trust or improves training. [4]
Its contribution here is narrower. On the measure used in that game, disclosure didn’t produce the drop the researchers expected. That’s a reason to question an assumption about participation, not a reason to stop asking about trust.
04
EVALUATION
How Should We Evaluate AI Avatars in Training?
Evaluate an AI avatar by whether it communicates accurately and supports what learners need to understand and do. A designer’s discomfort is a reason to investigate, not proof of instructional harm. Check learning and application, and assess trust separately.
For a particular course, I’d start by asking what the presenter is supposed to contribute. If I can’t explain why a person needs to be visible, human or synthetic, I’d reconsider the format before spending time improving the face.
Then I’d check the material itself. Is the explanation accurate? Can the learner hear it clearly and follow the demonstration? Are the captions right? An AI label doesn’t excuse problems I’d reject in any other production.
For the less obvious concerns, I’d test a short segment with people representative of the intended learners. I’d ask them to explain the key point or use it in a relevant task before drawing their attention to the visual detail that bothered me. Then I’d ask what they noticed, what distracted them, and whether anything made the information harder to trust.
Those answers serve different purposes. Someone noticing an odd smile doesn’t tell me whether they misunderstood the explanation. Someone answering a question correctly doesn’t tell me whether they trust the source enough to use the guidance. I need to avoid letting one answer stand in for another.
A small tryout would help me decide what to investigate next. It wouldn’t prove that the avatar had no effect. Where the choice matters enough, I’d compare versions that keep the instruction the same and change the feature in question. The amount of testing should reflect the consequences of getting the decision wrong.
That’s the approach I’d use, not a procedure these studies have validated for us.
05
CONCLUSION
What I’ve Landed On
I’m not abandoning AI avatars. I’m keeping them available without assuming that every example is suitable. Sometimes the right response will be to fix the visual. Sometimes it will be to use a different presenter, or no presenter at all.
If an avatar saves production time, I’d rather see that time go into stronger practice, assessment, and feedback. That’s the principle behind efficiency paying for excellence. The saving matters when it helps us make the learning better, not simply when it makes the production cheaper.
My eye will keep catching things. That’s still part of the job. But catching a flaw is not the same as proving harm, and I’ve caught myself letting the first stand in for the second.
I still want to fix the hand. I also want to know whether the learner can do what the training was meant to prepare them to do. My reaction shouldn’t answer that question for them.
¶
SOURCES
References
1. Anderson, C. (2016). TED’s secret to great public speaking [Video]. TED Conferences. Watch the talk.
2. Gu, C., Peng, Y., Nastase, S. A., Mayer, R. E., & Li, P. (2024). Onscreen presence of instructors in video lectures affects learners’ neural synchrony and visual attention during multimedia learning. Proceedings of the National Academy of Sciences, 121(12), e2309054121. https://doi.org/10.1073/pnas.2309054121. Open-access full text.
3. Mori, M. (2012). The Uncanny Valley (K. F. MacDorman & N. Kageki, Trans.). IEEE Robotics & Automation Magazine, 19(2), 98-100. Original work published in 1970. https://doi.org/10.1109/MRA.2012.2192811. Authorized translation at IEEE Spectrum.
4. Visser, B., van der Putten, P., & Zohrehvand, A. (2025). The impact of AI avatar appearance and disclosure on user motivation. In C. Anutariya, M. M. Bonsangue, E. Budhiarti-Nababan, & O. S. Sitompul (Eds.), Data Science and Artificial Intelligence: DSAI 2024 (Communications in Computer and Information Science, Vol. 2318, pp. 142-155). Springer. https://doi.org/10.1007/978-981-97-9793-6_10. First published online November 9, 2024; the publisher’s recommended citation uses 2025. The numerical details discussed here come from the authors’ 2024 preprint, The Impacts of AI Avatar Appearance and Disclosure on User Motivation.
In this post
