Make an AI Avatar Explainer With Consent, Label and Approval

The short answer: if a stock presenter will do and the video stays internal, an AI avatar explainer is an afternoon on Synthesia's Starter plan at $19 billed monthly, and the minutes hardly register, since every Synthesia plan comes with 10 minutes of video each month (read on the pricing page on 26 October 2026). If the presenter is a colleague, the work moves to paper that comes before the camera: a written consent with an end date, a script the colleague approves, and a visible AI label on every version. For example, a 4-minute explainer refreshed once a quarter uses 4 of the 10 included minutes in the months it changes, so the plan is not where the risk sits.

Disclosure: I may earn a commission through some links on this site. It does not change what this page says.

An avatar video needs three separate yeses, and the vendors collect only the first, once. The person whose face and voice are used agrees that the avatar may exist; someone agrees that this script may come out of that face; the viewer is told what they are watching. Single-tool screens are in how to use Synthesia and how to use HeyGen.

Before anyone records:

1. The arithmetic, so cost does not drive the choice. Synthesia meters about 100 credits per generated minute and includes the first 10 minutes on every plan (pricing page). For example, twelve 5-minute modules each generated twice come to 12 x 5 x 2 = 120 minutes a year, 10 a month on average: inside the included allowance.
2. The presenter you already have. A stock avatar needs no consent from your side, within Synthesia's rules: no promoted, boosted or paid social ads, no opinions or experiences put in the actor's mouth, and no standing in for a real employee without disclosure (platform integrity). Actors can withdraw too; Synthesia then retires the avatar, and later edits need a new presenter (retired avatars).
3. A consent record you can produce. For a personal avatar: who, which videos and topics, which channels and languages, the end date, how to withdraw, what happens to published videos then, and the stock alternative offered.
4. The number to read afterwards. The completion rate of whatever this video replaces.

What the law asks, read on 26 October 2026

A recording of a colleague's face and voice is information about an identifiable person, which is how Article 4(1) of the GDPR defines personal data (GDPR). If consent is your basis, Article 7 says you must be able to demonstrate it, and the person may withdraw it at any time, as easily as they gave it. Employees are the hard case: the EDPB's consent guidelines accept an employee's consent as free only when saying no has no adverse consequence at all, and their example is a film shoot where staff who decline get an equivalent desk elsewhere (EDPB guidelines 05/2020, paragraphs 21 to 23). The avatar version of that desk is a stock presenter, offered in writing as an equal option.

National law adds rights over a person's image. In the Netherlands, Article 20 of the Auteurswet bars publishing a portrait made on the commission of the person portrayed without their consent, unless agreed otherwise, and Article 21 lets a reasonable interest of the person portrayed block publication of any other portrait (Auteurswet). How that reaches an avatar is a question for counsel; a written agreement answers it in advance.

Article 3(60) of the AI Act calls generated or manipulated images, audio or video a deep fake when they resemble existing persons, objects, places, entities or events and would wrongly look authentic or truthful to a person. Article 50(4) makes the deployer, the organisation using the AI system (here, yours), disclose that the content is artificially generated or manipulated, and Article 50(5) asks for that clearly, at the latest at first exposure. Those duties apply from 2 August 2026 under Article 113. The amendment of July 2026, Regulation (EU) 2026/1744, left them as they were and gave providers of systems already on the market until 2 December 2026 for the machine-readable marking in Article 50(2) (consolidated AI Act). A breach of Article 50 can cost up to €15,000,000 or 3% of worldwide annual turnover, whichever is higher, and for an SME whichever is lower (Article 99(4) and (6)). This is the text as published, not legal advice.

The workflow, from release note to published video

1. Trigger: something changed. A release, a price, a policy. The source is a document the product owner has signed off, and the video says nothing it does not.
2. Consent before footage, for a personal avatar. Your written record comes first, then the vendor's check. In Synthesia, the person represented records the consent footage live, with an on-screen passcode; uploaded consent is refused (avatar from video). In HeyGen, every video-based digital twin needs a consent video from its subject, submitted with the footage, by webcam or by phone through a QR code (consent article); through the API or MCP server, create_avatar_consent returns a link valid for 24 hours, and you poll consent_status until it reads accepted (avatar consent). That page is plain about photo avatars: HeyGen runs no consent step and keeps no record for them, so the agreement is yours to collect.
3. The agent drafts. In Claude, add Synthesia as a custom connector at https://mcp.synthesia.io/mcp, paste the signed-off source and ask for a draft (Synthesia MCP). The connector, in public beta, has no avatar picker yet: the presenter comes from the Style or brand kit chosen in its setup widget, so keep cleared avatars in a Style of their own. Instructions for the agent: no claim beyond the source, no opinion or experience attributed to the presenter, and an opening line that says the presenter is AI-generated.
4. Two approvals before anything generates. Open the draft in the Synthesia editor. The product owner reads every scene against the source; a colleague who is the presenter reads it too, because the words come out of their face. Synthesia's avatar policy names these roles Talent, Owner, User and Approver, and asks for script standards that keep avatars from misrepresenting the people they show (avatar policy). Generate once both say yes.
5. The label. In Synthesia, add the AI label per scene from Media, AI Labels, or switch on Automatically add the AI label to my videos under Preferences; the workspace-wide switch needs an Enterprise admin (AI label). Synthesia also embeds C2PA provenance on its own, which viewers cannot see and which therefore does not replace the visible label (Article 50 guide). HeyGen commits to an invisible watermark, provenance metadata and a verification check by 2 December 2026 and leaves the visible label to you (HeyGen and the AI Act), so add a text line to its opening scene. Repeat it beside the player.
6. Publish, and write the register row. One row per published video: presenter, consent record, both approvals with dates, label present, where it runs, and a review date.
7. The end date arrives, or consent is withdrawn. Delete the avatar, then take down or regenerate every video the register lists for it. In Synthesia, deletion completes within 90 days with written confirmation (deletion article). In HeyGen, delete_avatar_group removes the group and all its looks (MCP tools).

The free version first

The free route avoids the question entirely: narrate your own slides or a screen recording. No likeness, no consent record, no deep fake. If a presenter matters, Synthesia's Basic plan costs nothing, shares by link and carries a Synthesia watermark (pricing page). The consent template and the register are a document and a spreadsheet, worth keeping on any plan.

Paying adds clean downloads, personal avatars and guest reviewers: Starter allows 3 personal avatars and 3 guests, Pro 5 of each (pricing page).

What each tool costs

All in US dollars, read on the vendors' pages on 26 October 2026.

Choose HeyGen when a founder or colleague has to speak many languages, since its avatars and translation share one account. Choose Synthesia when the label, the provenance and an avatar policy should come built in.

The default to override

Synthesia keeps an avatar in the workspace after its person leaves, lets everyone it was shared with keep making videos with it, and leaves videos already generated in place after the avatar is deleted (avatar policy). An avatar that stays usable keeps producing the generated minutes the plan meters. The date that stops it has to come from your register, because nothing in the default will.

Checking the avatar earned its place

For an explainer, read the completion rate and the click after it; HeyGen's share page reports views, average watch time and completion rate (sharing article). Set them against the version the avatar replaces.

Then a fair test: put the avatar video and a narrated-slides version of the same script on the same page for four weeks, swapping which one shows each week, and compare completion and the next click. A version without the label is not a variant to test on EU viewers.

The other measure is the register: every published avatar video with a consent record, both approvals and a label. My recommendation: audit it each quarter, and pause new videos the day one is missing.

FAQ

Does a stock avatar need the AI label too?

Treat it as yes. Synthesia's own guidance says a realistic presenter can fall within the disclosure rules even when the avatar is not an identifiable real person (Article 50 guide). The label costs a corner of the frame.

Can a colleague withdraw after the video is out?

Yes. Under Article 7(3) of the GDPR, withdrawal leaves earlier processing lawful but stops what comes next (GDPR). Delete the avatar, then use the register to take down or replace its videos, because deleting an avatar leaves generated videos in place.

Is the vendor's consent video enough?

No. It proves the person agreed to the avatar being made. HeyGen's moderation policy makes you responsible for having the actor's consent for each use or sharing as well (moderation policy), and Article 7(1) of the GDPR puts proof of consent on you. HeyGen's enterprise tier can even skip consent collection for accounts that sign an indemnity (avatar consent), which shows whose risk it is.