AI avatar platforms such as HeyGen and Synthesia turn a script into a video of a digital presenter speaking it, in dozens of languages, without a camera crew. The main risks are not technical. They concern consent from the person whose likeness is used, disclosure to viewers, accuracy of what the avatar says and how audiences react to synthetic presenters. Setting governance rules before the first video saves painful corrections later.
What do HeyGen and Synthesia actually do?
Both tools offer a library of stock avatars, based on paid actors who agreed to have their likeness used, and the option to create a custom avatar of a real person from recorded footage. You type or upload a script, pick a voice and language, and the platform renders a video with matched lip movement. Both support translation of existing videos and voice cloning, and both integrate with presentation and learning tools.
Synthesia has focused heavily on corporate training and internal communication. HeyGen is widely used for marketing, social content and video translation. Feature sets overlap and change often, so check current plans before committing.
What are the main risks of using AI avatars?
Consent and likeness rights
A custom avatar copies a real person’s face and voice. That person must give informed, written consent covering how long the avatar may be used, for which purposes and what happens when they leave the company. Both major vendors require consent recordings for custom avatars, but the vendor’s process does not replace your own contract with the employee or actor.
Disclosure
Viewers increasingly expect to know when a presenter is synthetic. The EU AI Act includes transparency duties for AI-generated or manipulated content that resembles real people, and platforms such as YouTube, TikTok and Meta ask creators to label realistic synthetic media. Hiding the fact risks trust and, in some contexts, rules.
Accuracy and claims
An avatar says exactly what the script says, with full confidence. If a script contains an error, the avatar delivers it convincingly in fifteen languages at once. Translated scripts need native review; machine translation can turn a careful claim into an overstated one.
Audience response
Some audiences accept avatars for training or product walkthroughs but react badly to them in emotional or high-trust contexts, such as healthcare advice, apologies or founder messages. Testing with a small audience before wide use helps.
Misuse and impersonation
A cloned executive voice and face is a security asset. Weak account security could let someone generate videos that appear to come from your CEO.
What are the technical limits?
Avatars have improved, but limits remain. Hand gestures and body movement are often stiff or repetitive. Expressions can lag behind emotional content. Pronunciation of brand names, Turkish characters or industry terms may need phonetic spelling. Long videos expose the repetition more than short clips. And the presenter usually stays in one framed shot, so editors add B-roll, screen recordings or graphics to keep attention.
Governance rules to set before the first video
A one-page policy covering these points handles most situations:
- Consent: written agreement for every custom avatar, with purpose, duration, revocation terms and payment where relevant.
- Allowed uses: list of approved content types (training, product demos, FAQs) and forbidden ones (crisis statements, financial or medical advice, anything appearing to be a live address).
- Disclosure: a visible label or spoken line stating the presenter is AI-generated, plus platform labels where offered.
- Script approval: subject-matter review for accuracy and legal review for regulated claims.
- Translation review: a native speaker checks every language version before release.
- Access control: two-factor authentication on the vendor account and a short list of people allowed to generate videos with custom avatars.
- Retention: when an employee leaves, their avatar is retired within a set period unless they agree otherwise.
- Record keeping: script, version, approver and publish date stored for every video.
How Alien Road applies AI avatars in video production
In a typical engagement the team first sorts the client’s video needs into those that suit avatars and those that need real footage. Product explainers, onboarding modules and multilingual FAQ videos usually fit; brand films and testimonials usually do not.
Discovery also covers the consent question. If a custom avatar of a client spokesperson is planned, the client’s legal team signs off the consent agreement before recording. The recording session itself follows the vendor’s guidance on lighting, framing and neutral expression, because those details decide how natural the avatar looks.
Production pairs avatar segments with screen recordings, motion graphics and captions. Scripts are written for spoken delivery, with short sentences and phonetic notes for difficult terms. Each language version gets native review. After launch, reporting compares completion rate and click-through for avatar videos against any earlier filmed versions, and comments are monitored for audience reaction to the synthetic presenter.
How do you decide whether avatars are right for a project?
Ask three questions. Will the content need frequent updates? Avatars make re-rendering after a script change cheap. Does it need many languages? Translation is where avatars save the most time. Does the message depend on genuine human presence? If yes, film a person.
Content that passes the first two and fails the third is the natural fit. Content that fails the first two rarely justifies the policy work.
Budget is a fourth consideration. Avatar platforms charge by plan and by minutes of video, and custom avatars usually sit on higher tiers. Compare that cost with what a half-day studio shoot would cost for the same set of videos, including the updates you expect over a year. For a library of short training clips in six languages the maths often favours avatars; for two hero videos a year it rarely does.
Where to start this week
List the videos your team updates most often and the languages each needs. Pick one low-stakes item, such as a software walkthrough, and produce it with a stock avatar and a clear AI label. Before you build a custom avatar of anyone, draft the consent agreement and the one-page policy. Getting those right first is far easier than retiring a video that should not have been made.
Keep reading: Video Production · AI Marketing