Lesson 6 of 8 · 7 min read

Consent and disclosure

What EU and US law say about AI likenesses of real people, a written consent checklist, and a disclosure line every viewer can see or hear.

This lesson is not legal advice

This lesson is not legal advice. It summarizes laws and vendor rules as of October 15, 2026, with sources. Ask a lawyer in your jurisdiction before you make an avatar or voice clone of a real person. Laws differ by country and by US state.

Each point below carries a label. “The text” means a law or policy says it in words. “Inference” means Emeri’s research drew a conclusion from the text that no regulator or court has confirmed. Treat inferences as questions to take to counsel.

The EU rule: disclose at first exposure

The text: Article 50(4) of the EU AI Act covers deployers who generate or manipulate image, audio or video that is a deep fake. A deep fake resembles existing persons, objects, places or events and would falsely appear authentic. The deployer must disclose that the content was artificially generated or manipulated.

The text: Article 50(5) says the disclosure must be clear and distinguishable. It must come at the latest at the first exposure. It must also meet the applicable accessibility requirements. These duties apply from August 2, 2026.

The EU’s Digital Omnibus delayed other parts of the Act. A Jones Walker analysis, dated July 16, 2026, says most Article 50 deadlines remain August 2, 2026. Only a narrow grace period, for machine-readable marking by tool providers, runs to December 2, 2026.

Inference: an avatar built from a real employee, saying words the employee never said, likely meets the deep-fake definition. A stock voice or avatar that resembles no real person may fall outside Article 50(4). The tool provider still has to mark it under Article 50(2). The Commission’s Article 50 guidelines are the place to confirm both points.

US rules on digital replicas and biometrics

The text: California’s AB 2602 (Labor Code §927) and New York’s General Obligations Law §5-302 took effect on January 1, 2025. Each makes a digital-replica clause unenforceable when three conditions all hold. The replica does work the person could have done in person. The clause lacks a reasonably specific description of the uses. The person had no lawyer or union in the deal. Illinois’s Digital Voice and Likeness Protection Act sets the same test for agreements made from August 9, 2024.

Inference: a one-line “we may use your likeness” clause in an employment contract is unlikely to hold for an employee without a lawyer or union. That applies in California, New York and Illinois. A separate, specific release is safer.

The text: Illinois’s Biometric Information Privacy Act lists voiceprints and scans of face geometry as biometric identifiers. A company must give notice, state the purpose and term, and get a written release before it collects them. Damages run up to $1,000 per negligent violation and $5,000 per intentional or reckless one. A 2024 amendment counts repeated collection by the same method as one violation per person.

GDPR: an employee’s consent is weak

The text: under the GDPR, biometric data used to identify a person is special-category data. In 2017, the EU’s data-protection regulators said consent from employees is unlikely to be freely given, because of the power imbalance at work. Employees must be able to withdraw consent without any penalty.

CSO Online’s summary of that position adds a practical point. An employer that relies on consent must offer an employee who objects an alternative, with no penalty. For a training video, the alternative is easy: a voice-over or a stock avatar.

Inference: in the EU, do not rely on an employee’s signature alone. Offer the alternative in writing, and make clear that saying no changes nothing about the person’s job.

What the vendors require

The text: Synthesia requires the person to record consent live for a personal avatar. A consent video cannot be uploaded, and the person in it must be the person in the footage. On Enterprise plans, a colleague can be sent a link to record consent, and the link expires after 7 days. If that colleague then signs up with the same email address, they become the avatar’s owner. Owners can delete their own avatars.

The text: HeyGen’s moderation policy requires the explicit consent of the person depicted, including for each use and sharing of the avatar. The person can ask for removal of their likeness at any time. The policy does not mention voice cloning. Inference: get separate written consent for a cloned voice.

The consent checklist and the disclosure

Emeri’s research built a five-item checklist from these texts. It is an inference, not a legal standard. First, a separate written release that names the uses, the channels, the duration and any use after the person leaves. Second, a withdrawal route with deletion steps. Third, an inventory of every video that uses the avatar or voice.

Fourth, an AI disclosure at first exposure. Fifth, the vendor’s own consent recording, made by the person. Add an offboarding step as well. An avatar can stay usable in a shared workspace after the person leaves, and deleting it does not delete videos already made.

Inference from Article 50(5): a visual watermark alone does not meet the accessibility requirement, because a blind viewer cannot see it. A spoken and captioned statement at the start reaches sighted and blind viewers. Put the same line in the course description in your LMS, as EdzLMS also advises.

Worked example · Larkfield Supply

What an avatar of Morgan would have needed

Lesson 1 turned down an avatar of Morgan, the warehouse manager. Riley lists what that avatar would have needed, so the next request gets a quick, honest answer. Riley then writes the disclosure for the AI voice the video does use.

  1. A separate release for Morgan, not a clause in Morgan’s employment contract. It names the returns video, the LMS as the only channel, a review date two years out, and deletion when Morgan leaves Larkfield.That answers the question you guessed at: Guess: is a likeness clause in an employment contract enough consent for an avatar of a colleague?
  2. A written alternative with no penalty. If Morgan says no, the video uses the AI voice with no avatar, and nothing else changes.
    What would you do next?Morgan says yes in the corridor. Is the consent done?Think it through, then show the answer

    No. A spoken yes is not a written release, and it names no uses or channels. Morgan would also have to record the vendor’s consent video in person. Riley cannot record it for Morgan.

  3. A withdrawal route and an inventory. Riley would add a line to the training log for every video that used the avatar, so one request could remove them all.
  4. A question for counsel. Riley does not know which state laws apply to Larkfield’s staff, so Riley writes the question down instead of guessing.
  5. The disclosure for the video Riley did make. The first words of each segment are: “This video uses an AI-generated voice. Larkfield’s training team wrote and checked the script.” The line takes about five seconds, inside the 10-second disclosure row from Lesson 4.
    What would you do next?The AI voice resembles no real person. Why disclose it at all?Think it through, then show the answer

    Article 50(4) may not require it, on the inference above. Riley discloses anyway, because it costs five seconds and learners notice. In one reported case, a whole training group said at once that they already knew the narrator was an AI voice.

The avatar of Morgan would have needed a release, a written alternative, an inventory, a vendor recording and a legal question. The AI voice needed one line, spoken at the start and shown in the captions and the course description. Riley keeps the checklist in the video plan for the next request.

Starter prompts

  • Draft a likeness release for counsel

    Draft a separate written release for using [person's name]'s [face / voice / both] in AI-generated training video. Include: the specific videos and uses [list them]; the channels where the videos appear [list them]; the duration and a review date [date]; what happens when the person leaves the company; how the person withdraws consent and what gets deleted; and a statement that saying no carries no penalty, with this alternative: [describe it]. Mark every place where local law may change the wording with [ASK COUNSEL]. Do not state that the release complies with any law. This draft is for a lawyer to review.

  • Write the disclosure line

    Write a disclosure line for the start of a training video. The video uses [an AI voice / an AI avatar / generated footage], and [who] wrote and checked the script. The line must be one or two sentences, under 10 seconds when spoken, and plain enough to put in captions and in the course description. Give three versions. Do not say the content is approved or accurate unless I tell you who approved it: [name the approver, or write "no approval"].

When not to use AI

  • Finalizing a release without a lawyer. An assistant can list the points to cover, but it cannot tell you what holds in your jurisdiction.
  • Using a colleague’s face or voice because they agreed in conversation. Get the written release and the vendor’s consent recording first.
Recall from memoryName the five items on the consent checklist for an avatar or voice clone of a real person.Answer it without looking back, then show the answer

A separate written release that names the uses, channels, duration and any use after the person leaves. A withdrawal route with deletion steps. An inventory of every video that uses the likeness. An AI disclosure at first exposure. The vendor’s own consent recording, made by the person.

Exercise · your project

Add a consent and disclosure section to your video plan.

  1. List every real person whose face or voice appears, or could appear, in your video.
  2. For each person, work through the five checklist items and the offboarding step, and mark what is missing.
  3. Write down your questions for counsel, including the countries and states where your learners work.
  4. Write your disclosure line, and note where it appears: spoken, in the captions and in the course description.

Sources