The session that argues against its own subject
Every other session in this course teaches you how to do something well. This one asks whether the thing is worth doing, using the numbers rather than the vibe. Two findings run through the whole session: the face is the weakest lever you have, and disclosure has never been shown to help you. Both are inconvenient. Both are what the literature says.
★ What you walk out with tonight. Three effect sizes you can quote from memory, one study design clean enough to survive a sceptical CFO, a genuine scientific disagreement you can describe without picking a side, and the one wording lever in this literature that is actually actionable.
LiveWhy a course about avatars spends a session attacking them3 min▶
Because you are going to be asked to defend a budget line, and the person asking will have read something. If the only research you know is the research your vendor sent you, you will lose that conversation, and you will deserve to.
There is a second reason, and it matters more. A course that teaches you to make synthetic presenters and then quietly skips the evidence on whether they work is doing the thing this course spends six sessions warning you about. So the awkward numbers go on the page.
You will see two different kinds of claim tonight and they are not equally strong. Mayer's effect sizes and the 108-student study are well-evidenced and consistent with each other. The disclosure literature is contested, with a large systematic review disagreeing with several large individual experiments. Anywhere the evidence does not agree, this page says so rather than picking the more dramatic side.
Self-studyThe 80 percent bar, and what stays with the vendors2 min▶
This session teaches roughly 80 percent of the working content of the research it maps to. What it does not do is turn you into someone who can appraise an education-research paper from scratch, and it does not replace reading the studies. Every number on this page traces to a source listed at the bottom.
If you are going to put any of these figures in a document that other people act on, open the source first. That is not a disclaimer, it is the standard the rest of this course holds vendors to, and it applies here too.
The face is the weakest lever you have
7 min liveRichard Mayer's multimedia principles come with median effect sizes, and three of them bear directly on whether to put a synthetic presenter in front of learners. Line them up and the ranking is not the one the product category would like.
LiveWhat each of those three levers actually is5 min▶
These are not vague quality ratings. Each principle is a specific, controlled manipulation, and each one is something you can decide on Monday.
| Principle | The manipulation | Median effect |
|---|---|---|
| Personalization | Conversational script register rather than formal register. "You" and "your" rather than "the learner" and "one". | d = 1.11 |
| Voice | Human narration rather than machine-synthesised narration. | d = 0.78 |
| Image | Adding the speaker's image on screen at all. | d = 0.22, which Mayer describes as "small-to-negligible" |
Read the ranking rather than the individual numbers. The script register is worth roughly five times the on-screen face. If you are spending a budget, the cheapest thing in that table is also the biggest, and it is the one nobody sells you.
The uncomfortable implication, stated plainly. A photoreal avatar with a formal, corporate script is optimising the weakest lever while ignoring the strongest. If your synthetic presenter says "learners should now consider the following process", you have spent money on d = 0.22 and thrown away d = 1.11 for free.
Self-studyThe voice penalty has largely dissolved, which changes the arithmetic4 min▶
The voice principle at d = 0.78 looks like a strong argument against synthetic narration. It was, when the comparison was human voice against traditional text-to-speech. That comparison is now out of date.
Dincer 2022, N = 51, compared human narration, modern text-to-speech and traditional text-to-speech. Human voice scored d = 0.90 and modern TTS scored d = 0.75. Both beat traditional TTS comfortably. But between human voice and modern machine voice there was no significant difference (p = .200), and cognitive load was statistically identical (p = .964).
What that does to the ranking. Voice was the middle lever, and modern synthesis has largely closed the gap. So the argument shifts: the strongest remaining lever is still the script register, and the weakest is still the face. The one that got cheaper is the one in the middle.
Craig and Schroeder reported the same reversal independently, which is why this is stated with more confidence than a single N = 51 study would justify on its own.
Engagement drops. Measured learning does not.
7 min liveIf you remember one study from this course, make it this one. It is the cleanest test yet published of an AI-generated instructor against a human instructor in a real course, and it splits the outcome in two directions at once.
LiveThe design, and why it is the cleanest one available5 min▶
Arkun-Kocadere and Caglar Ozhan, 2024, in IRRODL. 108 undergraduates: 48 taught by an AI-generated instructor, 52 by a human instructor, with 8 more in a focus group who saw both. Two 10-minute videos over two weeks. Video Engagement Scale as a post-test, Academic Performance Test before and after.
Two outcome measures, and they moved differently. Verbatim from the paper:
That pair of sentences is the single most useful thing in this session. It does not say avatars are fine and it does not say avatars fail. It says the thing people feel and the thing you can measure came apart, and your decision depends on which one your content is actually being judged on.
It also lines up with Mayer rather than contradicting him. The image principle sits at d = 0.22. A study finding no significant learning difference between instructor types is exactly what a small-to-negligible image effect predicts. Two independent lines of evidence pointing the same way is worth more than either alone.
LiveWhat the focus group said, and the escape hatch inside it4 min▶
The eight students who saw both instructors were interviewed. Their language is blunt, and it is worth quoting rather than paraphrasing.
Then comes the qualifier, and it is the most practically useful sentence in the paper:
Read that as a design instruction. The discomfort is real and it is conditional on the content being weak. Motivated learners on an interesting topic reported that they could set it aside. Unmotivated learners on a dull topic have nothing to set it aside for, and the discomfort is all that is left.
Which tells you where a synthetic presenter is least defensible: mandatory compliance training that nobody wants to be in. That is precisely the use case the category is most often sold for.
Self-studyThe corroborating meta-analysis, and its modest number3 min▶
Castro-Alonso, Wong, Adesope and Paas, 2021, pooled 32 effect sizes across N = 2,104 on learning with a pedagogical agent present versus absent. Overall g+ = 0.20.
Small, but positive. That is the honest characterisation, and both halves matter. Positive means an on-screen agent is not harmful on average. Small means it is not where your leverage is, which is the same conclusion the image principle reaches by a different route.
A note on reading effect sizes in front of executives. A g+ of 0.20 is a real effect and a weak one. If someone presents it as evidence that avatars improve learning, they are technically correct and practically misleading. The useful framing is comparative: 0.20 against 1.11 for the script register tells you where to spend, and no single number on its own does.
Teaching and persuading are different jobs, and the evidence splits
5 min liveThere is one direct head-to-head test of AI presenter video against human presenter video, and it found nothing. Before you use that as a green light, learn what a null result at that sample size is and is not worth.
LiveThe direct test, and the caveat that has to travel with it4 min▶
Leiker, Gyllen, Eldesouky and Cukurova, 2023, N = 83. One group watched an AI presenter video, the other a human presenter video. Both groups improved significantly (p < .001), with no significant difference in gains (p = .80) and no difference in how learners perceived the videos.
That is a genuinely useful finding. It is also weaker than it looks, and this is the part that gets dropped when the study is quoted in a deck.
Teach the caveat with the finding, every time. N = 83 is small. A null result at that size is weak evidence of equivalence, not proof of it. A study that small can fail to detect a real difference simply because it lacks the power to see one. No large replicated randomised controlled trial on this question exists. "No difference was found" and "there is no difference" are different sentences, and only the first one is supported.
The correct way to carry this into a decision: the evidence does not currently support paying much for a face, and it also does not license you to claim equivalence has been demonstrated. Both halves, or neither.
LiveAnd one study pointing firmly the other way4 min▶
Everything so far has been about learning. Change the outcome measure to persuasion and the picture changes with it.
Chen, Zeng and Qiu, 2025, ran three experiments with a combined N = 839 on policy adoption intention. A human anchor produced significantly higher adoption intention (M = 5.89) than a digital anchor (M = 4.21). That is not a small gap.
It is also not replicated, so hold it at the same standard as everything else on this page. But the direction is worth taking seriously, because it is the direction most organisations are actually operating in.
| The job | What the evidence supports | Strength |
|---|---|---|
| Teaching | The face is close to neutral. Do not pay much for it. Spend on script register instead. | Several converging lines: d = 0.22, g+ = 0.20, N = 108, N = 83 |
| Persuading | A human still wins, and by a wide margin on adoption intention. | One unreplicated set of three experiments, N = 839 |
The synthesis this course will actually defend. For teaching content, the evidence does not support paying much for a face. For persuading people, the limited evidence says a human still wins. Those are different jobs, and blurring them is how a vendor deck gets written. Your onboarding module and your CEO's strategy announcement are not the same decision.
Does disclosure cost you anything? Nobody agrees.
8 min liveThis is the part of the session where the honest answer is that the field has not settled it. Large individual experiments find a penalty for labelling content as AI. The largest synthesis of the literature finds no consistent penalty. This page teaches the disagreement rather than the more dramatic side of it.
LiveThe systematic review, which is the biggest thing in the room5 min▶
Licenji and Hoxha, 2026, in Frontiers in Artificial Intelligence, searched Scopus and Web of Science and read 47 full-text studies spanning 2018 to 2026. Their conclusion, verbatim:
They describe audience responses as "contingent, varying by design, context, and measurement target". Which is a careful way of saying that the answer depends on what you labelled, who you asked, and what you measured, and that anyone quoting a single clean number is over-claiming.
An earlier version of this course's own source map got this wrong. It stated that labelling suppresses trust as though it were settled. It is not, and the correction is recorded rather than quietly edited out. If a course about disclosure will not disclose its own corrections, it has no standing to lecture anyone.
LiveBut the same review finds the thing you should build on4 min▶
The review does not rescue disclosure. It removes the penalty argument and leaves something more awkward standing.
Nobody finds a benefit. Not one study across the whole synthesis. Whatever else is contested, that is not. Disclosure is not a persuasion tactic under any reading of this evidence, and any vendor or agency telling you that labelling content builds trust is asserting something no study supports.
So why disclose at all? For a reason none of this research touches. Disclosure removes the later moment where someone finds out. That is a durability decision and an ethical one, and it is unaffected by whether labelling costs you a few points of perceived accuracy today. Session 1's simulator made the same argument mechanically: disclosure is worth zero on the craft axis and everything on the survivability axis.
Self-studyThe individual experiments, which are large and real4 min▶
The review disagrees with these studies. It does not make them go away, and they are big.
Altay and Gilardi 2024, two pre-registered experiments, N = 4,976 across the US and UK. Participants did not equate "AI-generated" with "false". But labelling headlines as AI-generated lowered their perceived accuracy and participants' willingness to share them, regardless of whether the headlines were true or false.
Longoni, Fradkin, Cian and Pennycook 2022, N = 4,034. People rated AI-written headlines as less accurate, and were more likely to incorrectly rate headlines written by AI as inaccurate when they were actually true.
How to hold two disagreeing bodies of evidence at once. These are headline-credibility experiments. The review pooled 47 studies across many designs, contexts and outcome measures. Both can be accurate descriptions of what they measured. What you cannot do is pick whichever one supports the decision you already made, and that temptation is the reason this section exists.
Self-studyThe implied truth effect, which nobody plans for3 min▶
There is a second-order problem with labelling that only shows up at scale. If you label some of your content and not the rest, the unlabelled remainder does not stay neutral. It quietly gains credibility it has not earned.
That is the implied truth effect: attaching warnings to some items makes the untagged items look validated by comparison. Applied to AI provenance, partial labelling produces a two-tier ecosystem where "no label" starts to read as "verified human", which was never what the absence of a label meant.
The operational consequence. A disclosure policy that covers only your flagship content is not half a policy, it is a different policy with a side effect you did not choose. Decide the scope deliberately, and write down what an unlabelled item is supposed to mean.
The one lever that is actually actionable
★ 18 min · everyone writesEverything above tells you what not to claim. This part gives you the one finding in the whole disclosure literature you can act on tonight, and then makes you rewrite a sentence you were probably about to publish.
Find the disclosure sentence you already use. On a video, a landing page, a footer, a course module. If you cannot find one, write the sentence you would have used. That is your starting text.
Check which of the three words it leans on. A between-subjects experiment with n = 415 found message credibility varied significantly by whether the label said content was "influenced", "assisted" or "generated" by AI, with "AI-generated" faring worst. Source credibility was unaffected by the label entirely.
Rewrite it to describe what actually happened. If a human wrote the script and reviewed the output, "AI-assisted, human-reviewed" is both more accurate and better tested than "AI-generated". If nothing human touched it, you do not get to use the softer wording. Accuracy first, wording second.
Now delete the claim you were about to make. Search your draft for any sentence implying that disclosing builds trust, shows integrity to the audience, or increases credibility. No study in a 47-study review found a positive trust effect. Cut it.
Replace it with the argument that survives. Write one sentence saying why you disclose that does not depend on the audience rewarding you for it. If your sentence still needs the audience to approve, it is not the right sentence yet.
Last, decide the scope out loud. Which content gets labelled, and what does the absence of a label mean on your channels? If you have not decided, the implied truth effect decides for you.
The design decision worth naming. Every other lever in this session is something the evidence tells you not to bother with. The wording is the exception. It is free, it is tested, and it is the difference between describing a process and triggering an assumption of full automation with no human in it.
Before session 5
30 min- Open the last thing your organisation published with a synthetic presenter in it, and read the script aloud. Count the sentences written in a formal register that could have been conversational. That count is the d = 1.11 lever you left on the floor.
- Find one vendor claim about learning outcomes and check it against tonight's numbers. Vendor decks quote real studies. Check whether the caveat travelled with the finding, and whether a null result is being presented as proof of equivalence.
- Write the two-column table for your own use case. Left column: is this content teaching or persuading? Right column: what the evidence supports for that job. If a piece of content is doing both, split it.
- Rewrite your disclosure wording using the n = 415 finding. Then check the rewrite is still true. Softer wording that misdescribes what happened is worse than the harsher wording that is accurate.
- Optional, and it will change a meeting: take the sentence "disclosing builds trust with our audience" to whoever wrote it, and ask what evidence it rests on. Have the 47-study finding ready. It is a short conversation.
Three questions
Clear all three and session 4 gets stamped in your passport.
1 · Your team is deciding where to spend on a video series. What do Mayer's median effect sizes say about where the leverage sits?
Mayer describes the image principle at d = 0.22 as "small-to-negligible". The conversational script register is the strongest lever in the set and also the cheapest, which is why nobody sells it to you.
2 · In the 108-student study with an AI-generated instructor, what actually differed between the two groups?
Engagement drops, measured learning does not. The focus group added the qualifier that matters: students reported distraction, discomfort and disconnectedness, but said those feelings could be ignored when the topic was interesting or they were focused on learning.
3 · A colleague says "the research proves that labelling content as AI suppresses trust in it". What is the accurate version?
Altay and Gilardi at N = 4,976 and Longoni et al. at N = 4,034 found a penalty. Licenji and Hoxha's 47-study review found no consistent one and called responses "contingent". The one thing both sides agree on is that no study showed a clearly positive trust effect, so C is the only option the evidence rules out entirely.
What this session teaches, and where it came from
Every fact on this page, and its verification tier, is recorded in the course's source map., including the recorded correction to an earlier draft that overstated the disclosure finding.
Session 4 cheat sheet · pin this
Tonight was about what the evidence says when you are the one making the avatar. Session 5 turns it around: what happens when someone builds one of you. Three seconds of audio, a finance approval that was not real, and the organisations that lost money against the one that lost nothing and published why.