How this course works
Six sessions, 45 minutes each, and one decision you write down and can defend. Not a video. A documented decision about whether, how and on what terms your organisation uses a synthetic presenter, in a form you could hand to a lawyer, a works council, or the person whose face it is.
★ What you walk out with tonight. The actual consent architecture of the three major vendors, read from their own policy pages rather than their marketing. A test for whether a contract is safe to sign. And an honest answer to whether your audience will notice, which is no.
LiveWhy this course leads with the decision3 min▶
Making a synthetic presenter is now easy, cheap and fast. That is precisely why the interesting question moved. The constraint is no longer capability, it is judgment, and judgment is the part that gets audited afterwards.
So this course is organised around a decision rather than a workflow. Session 6 ends with a standing position your organisation can actually hold, rather than a video you made once.
McDonald's Netherlands published an entirely AI-generated Christmas advert and pulled it three days later after viewers called it AI slop. It was made with a named agency and professional AI specialists. Nothing about it failed technically. It failed on reception, which is not a problem you can fix with a better model.
Self-studyNames that changed, and will date any older material3 min▶
- HeyGen's "Interactive Avatar" no longer exists under that name. That URL now redirects to a separate brand, domain and subscription called LiveAvatar, and avatars are not cross-compatible between the two products.
- D-ID is alive, and is the acquirer rather than the acquired. It announced the acquisition of simpleshow in September 2025, with the combined company operating under the D-ID brand. Any list that has quietly dropped D-ID is wrong.
- D-ID renamed its avatar tiers. "Express Avatar" is now V3 Instant, "Premium+" is now V3 Pro, and photo avatars are V2. The old names survive in D-ID's own URLs, which tells you how recent this is.
None of this is trivia. A guide that still says "HeyGen Interactive Avatar" was written before this year, and in this field that is long enough to matter.
Whether consent is required depends on which button you press
8 min liveThe three major vendors give three genuinely different answers to "may I make an avatar of a real person", and the differences are not marketing noise. They are the whole risk, and almost nobody reads them before signing up.
LiveThe three sentences that should decide your vendor5 min▶
All three are quoted verbatim from the vendors' own pages, and each one is doing a lot of work.
Read that second one carefully. It is true of a generated face. It is not true of a photograph of your colleague, and nothing in the product distinguishes between the two.
What that last one actually tells you. The consent step is a liability-allocation mechanism as much as an ethical one, and at enterprise scale it is purchasable. That is not a scandal, it is how commercial risk works. But it does mean the protection you are relying on may not be the protection you think, and it is worth knowing which one you have bought.
Self-studyWhat the person actually says, when anyone publishes it4 min▶
Only one of the three publishes the words the person is asked to read. It matters, because a consent recording is only consent if it contains consent.
Provided in roughly thirty languages, to be read in the actor's native language. It says who they are, what is being made, and what it is for.
That is a performance-capture prompt. It contains no words of permission, authorisation or consent. HeyGen says consent is "validated semantically", and that may well be sufficient in practice, but the published artefact is not a consent statement. D-ID publishes no wording at all.
Self-studyWhat enterprise money buys, which is different at each vendor4 min▶
| Vendor | What the enterprise tier exempts you from |
|---|---|
| HeyGen | The consent step itself, against a signed indemnity agreement |
| Synthesia | Content restrictions. Political and news content is enterprise-only, and the impersonation prohibition is worded to apply to non-enterprise plans |
| D-ID | The automated public-figure block, which Advanced and Enterprise tiers can request an override for |
Three vendors, three different things behind the same paywall. If your procurement conversation is about minutes and resolution, it is about the wrong thing.
Three actors said yes, and here is what yes turned out to mean
6 min liveThese are named people, reported by a wire service, with the vendor on the record. They are the clearest available answer to what a consent form is worth when it is written badly.
LiveThe test a lawyer gave, and it fits in one line4 min▶
Alyssa Malchiodi, quoted in the same reporting, gave the red flag to look for in any likeness agreement:
Three words to find and interrogate in any contract in front of you: perpetual, irrevocable, and transferable. Each one individually is normal commercial language. Together, applied to a human face, they mean the person can never get it back.
And the sentence that should be read aloud to anyone about to consent. It comes from Synthesia's own user terms, and it is to the vendor's credit that they wrote it down: "Because Customer, and not you, will own any videos that incorporate those Custom Avatars ... please ensure you are fully informed by Customer about how it intends to use them before you consent."
The person in the video does not own the video. Most people saying yes have not understood that, and it is the single thing most worth telling them.
Your audience will not notice, and that settles the disclosure question
5 min liveThere is a tempting argument that disclosure is unnecessary because people can spot synthetic media anyway. It is worth taking seriously for exactly as long as it takes to look at the evidence.
LiveWhat this does and does not license you to conclude4 min▶
A 2024 meta-analysis pooled 56 papers, 137 effects and 86,155 participants on how well people detect synthetic media. Overall accuracy was 55.54 percent, with a confidence interval that crosses the 50 percent chance line. Audio was the best-detected modality at 62.08 percent and text the worst at 52.00 percent. Participants were noticeably better at recognising real material than synthetic material.
The conclusion that follows is narrow and firm: disclosure has to be a deliberate act by the publisher, because it cannot be outsourced to the viewer's perception. Every argument of the form "people can tell" fails on this evidence.
An honesty note about this number, because this course holds itself to the standard it teaches. The publisher's page and the repository copy were both inaccessible when this was compiled, so these figures come from the study's abstract and the author's own publication listing rather than from the full text. Treat them as reported rather than personally verified, and check the DOI before you quote them in a board paper.
Two axes, and only one of them is craft
★ 18 min · everyone buildsThe playground below drives a real viseme track computed from whatever you type, so the mouth genuinely follows your words. The two scores are a teaching model. What they encode is not: disclosure does not make a synthetic presenter more convincing, which is exactly why it is a separate decision from craft.
Type something and press speak. Watch the viseme track underneath. Commas and full stops become genuinely longer held pauses, because prosody is timing. Change the words and the track changes with them.
Turn the four craft levers on one at a time. Believability climbs from 24 to 84. All four are real production choices: pause at punctuation, blink on a human cadence, allow small head motion, match the mouth to the actual syllables.
Now switch on disclosure and watch believability not move at all. It stays at 84. That is the entire point, and it is why the lever is worth plus zero on the craft axis. Disclosure was never a craft decision.
Switch disclosure back off and press "now someone finds out". Believability holds at 84, because the avatar is exactly as good as it was a second ago. Trust falls to 6 percent. Everything it ever said is now in question, including the parts that were true.
Sit with the middle state for a moment. Undisclosed and nothing has gone wrong yet is where most organisations actually are. It looks identical to safety right up until it does not, and the liability was there the whole time.
The design decision worth naming. Disclosure is worth zero points on the believability axis in this simulator, deliberately. Anyone who tells you disclosure improves your content is selling something. It does not. It changes what your content is worth, which is a different and more durable thing.
Before session 2
30 min- Open the actual policy page of whichever vendor you are closest to using. Not the pricing page. The acceptable use policy, the terms, and any biometric notice. Read what consent it requires for the specific avatar type you would use.
- Find the three words. Search that contract for perpetual, irrevocable and transferable. Write down where each appears and what it applies to.
- Answer the ownership question in one sentence. If you make an avatar of a colleague, who owns the videos? Find the sentence that says so, and note whether your colleague would understand it.
- Write down what happens when they leave. Nobody publishes this. Your answer is the beginning of your own policy, and session 6 finishes it.
- Optional, and it changes how people think: show a colleague a synthetic presenter without telling them, then tell them. Watch which of the two reactions is stronger. That reaction is what you are managing.
Three questions
Clear all three and session 1 gets stamped in your passport.
1 · You want to make a talking avatar from a photograph of a colleague. Which vendors require a recorded consent step for that?
HeyGen states photo avatars "depict no real, identifiable person and do not require consent", which is true of a generated face and not of your colleague's photograph. D-ID's V2 avatars need only an image. Synthesia requires live consent and bars creating one for another person even with their permission.
2 · In the simulator, switching on disclosure left believability unchanged at 84 percent. Why is that the correct behaviour?
That separation is the point of having two axes. Craft levers move believability. Disclosure moves whether the trust you have is durable, and pressing "now someone finds out" shows what the undisclosed version was actually worth.
3 · A colleague argues that disclosure is unnecessary because people can spot AI presenters anyway. What does the evidence say?
Audio was the best-detected modality at 62 percent and text the worst at 52. Nothing in that data supports relying on the viewer. Disclosure has to be an act by the publisher.
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 a dated list of what is deliberately not stated.
Session 1 cheat sheet · pin this
Tonight you saw that the consent step varies by vendor, by avatar type, and by what you pay. Session 2 goes into the contract itself: what a consent form has to contain to be worth anything, what the three actors' contracts were missing, and how to write the clause about someone leaving that no vendor will write for you.