A learning technologist hits play on the finished onboarding video. A photoreal presenter in a navy blazer welcomes the new hire, voice warm and unhurried, and explains the company's code of conduct across nine clean minutes. Nobody filmed it. No human said those words. The presenter does not exist, the voice was cloned from a sixty-second sample, and the script was drafted by a model. The video looks broadcast-grade and ships to 4,000 employees on Monday. Then a works-council representative asks one question that nobody on the team prepared for: "Does the person watching this know it is synthetic, and where does it say so?" The silence that follows is the subject of this lesson.

Three Categories, Not One Magic Button

The phrase "AI video" hides at least three different technologies, each with a different cost curve, a different failure mode, and a different disclosure obligation. Conflate them and you will buy the wrong tool, trust the wrong output, and disclose the wrong thing. Separate them and you can place each one on a real production decision with eyes open. The three categories are synthetic avatars (a generated on-screen presenter), AI voice (synthetic or cloned narration), and generative video (AI-produced footage, b-roll, and animation). A single platform often does all three in one run, which is exactly why learning professionals stop seeing the three jobs and start seeing one magic button. Your first task is to un-blend the button.

A quick definition before we go further, because the vocabulary is load-bearing. Synthetic media means audio, video, or images generated or substantially altered by AI rather than captured from reality. Why you care: the moment a presenter, a voice, or a scene is synthetic rather than filmed, a new set of duties attaches to it, around disclosure, around accessibility, and around what your learners and your legal team are entitled to know. The technology is not the risk. The undisclosed, unverified, inaccessible use of it is.

Synthetic Avatars: The Presenter Who Was Never There

An avatar is an AI-generated on-screen human who delivers your script, lips synced to narration the model also produces. You type the words, pick a stock presenter or a custom likeness, and the platform renders a person speaking them. Why a learning team reaches for this: a polished talking-head video that once cost a studio day, a presenter's fee, and a week of editing now costs a text box and a render queue. Update a policy threshold and you retype one line instead of rebooking a shoot. The savings are real, and they are also the trap, because the same frictionlessness that lets you fix a typo in seconds lets a wrong policy threshold reach 4,000 people in seconds. Avatars are a generation technology wearing a human face, and generation's failure mode, the confident wrong claim, does not disappear because the messenger looks trustworthy. It gets worse, because a credible human face lends authority to whatever it says.

AI Voice: The Narrator Who Is a Clone

AI voice splits into two meaningfully different things. Synthetic voice is a generated narrator built from a model's training, not tied to any real person. Voice cloning is a synthetic copy of a specific, identifiable human voice, built from a sample as short as sixty seconds. The distinction matters enormously for consent and disclosure. A generic synthetic narrator raises a disclosure question. A clone of your CEO's voice, or a former employee's, or a voice actor whose contract never contemplated cloning, raises a consent question, a likeness-rights question, and a much sharper disclosure question. The bright line: you may not clone a real, identifiable voice without that person's documented, specific, informed consent for that use, full stop. "We had a recording of them" is not consent to synthesize new sentences they never spoke.

Generative Video: The Footage of a Place That Does Not Exist

Generative video produces moving footage from a prompt: b-roll of a warehouse, an animated process diagram, a scene of a customer interaction. It is the youngest and least predictable of the three, and its failure mode is not just the wrong claim but the wrong depiction. AI-generated footage of a "lab safety scene" can show a worker handling a chemical without gloves, an exit sign in the wrong place, or a procedure performed in the wrong order, and a learner absorbs the depicted behavior whether or not the narration is correct. In a safety or compliance context, a plausible but wrong visual is its own hallucination, and it ships at the same scale as the script.

The Trade-Offs That Actually Decide the Build

Once the three categories are separate, the real decision is not "should we use AI video" but "for this specific learning need, what does each category buy us and cost us." The honest answer is a trade-off table, not a sales pitch. Speed and update cost favor synthetic media heavily. Trust, nuance, and the felt presence of a real human favor a real human, and in some contexts that felt presence is the learning. A leader's authentic message about a layoff, a CEO's word on a values commitment, a frontline safety briefing where a learner needs to believe a real person stands behind the procedure: these are places where a synthetic presenter does not just raise a disclosure issue, it can quietly corrode the trust the learning depends on.

Read the table as a placement guide, not a verdict. AI video earns its place in high-volume, frequently updated, lower-stakes-of-depiction content, where the cost of reshooting a real human every quarter is the thing it actually solves. It earns suspicion the moment the learning leans on a real human's authenticity or a procedure's depicted correctness. The skill is knowing which row you are in before you pick the tool.

The Vendor Category Map: Orientation, Not Endorsement

You will be shown logos. Treat them as a map of categories, never as a verdict on quality or safety. The AI video, avatar, and voice space includes platforms positioned mainly as avatar and presenter studios (for example, Synthesia, HeyGen, Colossyan), platforms positioned around animation and explainer video (for example, Vyond), and voice and audio editing tools positioned around synthetic and cloned narration and podcast-style editing (for example, ElevenLabs, Descript). These names orient you to what a category does. They do not tell you whether a given video's claims are true, whether the build is accessible, or whether you have consent to clone a particular voice. A vendor's published speed and savings figures are vendor-reported by definition: a useful signal of what the category enables, never a load-bearing fact in your business case or your audit file.

The logo on the studio never owns the disclosure, the consent, or the accuracy. You do. The obligation does not transfer to the platform.

The reason to stay vendor-neutral is not diplomatic, it is operational. Tools change monthly; a feature that auto-adds a disclosure label this quarter may not next quarter, and a platform's default may flip from "watermark on" to "watermark off" in an update you did not read. If your disclosure practice depends on a vendor default, your disclosure practice is one release note away from failing. Build the practice so it holds no matter which logo renders the video.

Disclosure: What Your Learners and Your Legal Team Expect

Synthetic-media disclosure means telling the audience, clearly and accessibly, that a presenter, a voice, or a scene was generated or substantially altered by AI rather than captured from reality. Why you care: an undisclosed synthetic presenter in a regulated training, a works-council environment, or a jurisdiction with transparency rules is not a stylistic choice, it is a governance gap your legal team will find before the auditor does. The 2026 reality is that transparency obligations for AI-generated and manipulated media are tightening across jurisdictions, the EU AI Act among them, and the safe posture is to disclose by default rather than to litigate, per video, whether you were strictly required to. Disclosure is cheap. The argument about whether you needed it is not.

Disclosure done well is not a legal disclaimer buried in a terms link. It is a practice with four parts, and each part has to survive an accessibility check, because a disclosure a blind learner cannot perceive is not a disclosure.

• Visible on-screen notice. A persistent or clearly placed label that the presenter or footage is AI-generated, in the video itself, not only in a description that travels separately from the file.
• Spoken or captioned statement where it matters. In higher-stakes content, the disclosure is also in the narration and the captions, so it reaches a learner regardless of how they consume the video.
• Metadata and provenance. A record, ideally machine-readable, that the asset is synthetic, who generated it, from what source script, and when. This is the file an auditor reconstructs the history from.
• Consent records for any cloned likeness or voice. Documented, specific, informed consent for the exact use, retained where you can produce it on request.

Notice that disclosure and accessibility are the same discipline wearing two hats. A spoken disclosure with no caption fails a deaf learner. An on-screen label with no audio cue fails a blind learner. The synthetic-media disclosure and the WCAG 2.2 AA conformance target are not two separate checklists; they are one build practice. The lesson on accessibility from the first draft is the other half of this one.

The Consent Question Nobody Asked

Return to the cloned voice. A common, costly mistake is treating a voice actor's existing recording, or a former executive's archived keynote, as raw material for cloning. It is not. Consent to be recorded saying specific words is not consent to have a model generate new words in that voice forever. The same applies to a custom avatar built from an employee's likeness: when they leave, does their face keep delivering your compliance training? Decide the answer, in writing, before you build the likeness, not after the person has left and asked you to stop. The bright line stands: no cloned identifiable voice or likeness without documented, specific, informed consent for that use, and a defined retirement path when consent ends.

A Worked Example: Before and After

Return to the onboarding video from the opening, and watch two versions of the same build.

Before (the magic button). The team uploads a code-of-conduct script, picks a stock avatar and a confident synthetic voice, and renders nine minutes in an afternoon. It looks finished, so it ships. Three things were never decided. First, nobody verified that the avatar's spoken policy thresholds matched the actual approved policy; one figure was drafted by the model and never checked against the source. Second, nothing in the video discloses that the presenter is synthetic, because the platform's default watermark had been turned off in a prior project and nobody turned it back on. Third, the auto-generated captions were accepted unread, and they render the phrase "reasonable accommodation" as "reasonable accommodations" in one place and drop a "not" in another, quietly reversing a sentence. When the works-council representative asks the disclosure question, the team has no answer, no consent record for the voice, no provenance file, and a caption error in a legal definition. The video is pulled. The rebuild costs more than a real shoot would have.

After (the three jobs, disclosed and verified). The same team treats the build as separable jobs with gates. The script's every policy claim is checked against the approved code of conduct before a single frame renders, so the threshold is right. The avatar choice is a deliberate decision: for onboarding policy content, a synthetic presenter is acceptable, and a persistent on-screen label plus a one-line spoken disclosure in the opening make it transparent. The voice is a generic synthetic narrator, not a clone, so no consent record is needed, and that choice is logged with the reason. Captions are generated, then read and corrected against the script word for word, with the legal definitions verified character for character. A short provenance note records that the asset is synthetic, from what script, generated when, by whom. When the works-council representative asks the question, the lead answers in one breath: yes, it is disclosed on screen and in the narration, here is the provenance note, the voice is a non-cloned synthetic narrator, and the captions were human-verified against the approved script. Same tool, same speed, completely different fate, because the three jobs were named, the disclosure was built in, and the claims were verified.

The lesson is not that synthetic presenters are forbidden. It is that an undisclosed, unverified synthetic presenter is a liability shipped at scale, and a disclosed, verified, accessibly built one is a defensible, cost-saving choice. The video did not change. The accountability did.

Key Takeaways

• "AI video" hides three different technologies: synthetic avatars (a generated presenter), AI voice (synthetic or cloned narration), and generative video (AI-produced footage), each with its own cost curve, failure mode, and disclosure duty.
• Avatars are generation wearing a human face: a credible synthetic presenter lends authority to whatever it says, so every policy threshold and procedure it delivers must be verified against the source before it renders.
• Voice cloning of a real, identifiable person requires documented, specific, informed consent for that exact use, full stop; an existing recording is not consent to synthesize new sentences.
• Generative video can teach a wrong behavior through a wrong depiction, gloves off, exit sign misplaced, steps out of order, even when the narration is correct, so depicted procedures need the same verification as spoken claims.
• The build decision is a trade-off, not a default: synthetic media wins on speed and update cost in high-volume content, and loses where a real human's authenticity or a procedure's depicted correctness is the learning.
• Synthetic-media disclosure is a four-part practice: a visible on-screen notice, a spoken or captioned statement where it matters, machine-readable provenance, and consent records for any cloned likeness or voice.
• Disclosure and accessibility are one discipline: a spoken disclosure with no caption, or an on-screen label with no audio cue, fails a learner and fails the WCAG 2.2 AA target.
• Vendor logos are a category map, never an endorsement; the platform never owns the disclosure, the consent, or the accuracy, and vendor speed and savings figures are vendor-reported, never load-bearing.