Ethical AI Video: What to Check First

If you generate video of a person for anything commercial, the risk is not philosophical. It is a takedown, a platform ban, or a claim from the person whose face you used. The checks below are the ones that actually come up, in the order they tend to bite.

If the face belongs to a real person, and not a generated character, you need their permission for the specific use. Agreeing to be filmed is not the same as agreeing to be regenerated saying things they never said, and that distinction is where most disputes start.

Scope it explicitly. Which videos. Which languages. For how long. Whether the material can be edited later. Whether it can be used after they leave the company, which is the case teams forget until an employee resigns and their face is still fronting the onboarding video.

For your own face this is trivial. For an employee, a client, or a contractor, it is not.

A voice is treated separately from a face in most frameworks, and this applies equally whether you are dubbing existing footage or generating from scratch with an AI avatar, and several jurisdictions now protect it explicitly. Consent that covers likeness does not automatically cover voice. Voice carries its own body of law, and the copyright position on cloned voices is worth reading before you record anyone.

This matters more than it sounds, because voice cloning needs so little material. Twenty to thirty seconds of clean speech is enough to produce a usable clone. Any tool that clones from a short sample without asking whose voice it is has moved the entire compliance burden onto you.

Most consent people collect is too vague to rely on. A clause that holds up names five things:

None of this requires a lawyer to draft from scratch. It requires deciding the answers before you need them.

Ask whether the model was trained on licensed material. The Content Authenticity Initiative ↗ is where the provenance standards are being written. Our own model comparison covers what each platform discloses. Most vendors will not answer precisely. The ones that do are worth preferring, and the question is hardening quickly enough that a vague answer today may become a problem later. Provenance claims vary a great deal between products, and the model catalogue records which engine produced a given clip.

Watermarking marks the file. Disclosure tells the viewer. You may need one, both, or neither depending on where your audience is. The distinction matters commercially as well as ethically, which is why tools differ so much on watermarking and commercial use.

The EU AI Act ↗ requires that synthetic media be disclosed to viewers, with transparency obligations applying from August 2026. The full text and timeline ↗ are worth reading rather than taking second hand. If any meaningful part of your audience is in the EU, assume disclosure applies to you regardless of where your company sits.

These are moving quickly and they are not uniform. The pattern across most new rules is consistent even where the detail is not. Non consensual sexual imagery is criminal, and the FTC's business guidance ↗ is the plainest running commentary on where US enforcement is landing. Political synthetic media near an election is restricted or must be labelled. Commercial use of somebody's likeness without permission is actionable.

Check locally rather than assuming your jurisdiction resembles the one you read about.

Platform rules are usually stricter than law and change faster, which affects where you can publish avatar video. Most large platforms now require creators to label realistic synthetic media, and enforcement is increasingly automatic rather than complaint driven. YouTube ↗ and TikTok ↗ both publish their rules, and both changed them within the last year. Each platform words its rule differently, and the upload requirements for YouTube, TikTok and Instagram are the practical version of this section.

The practical consequence is that a video can be entirely legal and still be removed. Check the platform policy before you produce a campaign around a format, not after.

Can you remove a cloned likeness and voice, and does removal propagate to the vendor's systems rather than just hiding it in your dashboard.

If somebody withdraws consent you will need a mechanism, and the moment you need it is a bad moment to discover you do not have one. Ask before you are in that conversation.

The objection to disclosure is always that it undercuts the work. In practice the requirement is that a viewer can tell, not that the video apologises for itself. There is a longer treatment of the same tension in the creator guide to disclosing AI content.

What generally satisfies it: a line in the video description, the approach we describe for dubbed video, an on screen label in the first few seconds, or the platform's own synthetic media toggle where one exists. What generally does not: burying it in a linked terms page, or a watermark so faint it survives only in the source file.

The pattern worth copying is the one used for sponsored content. Nobody pretends a paid partnership label ruins a video, because it is short, early and unambiguous. Treat synthetic media the same way and the compliance question stops being a creative question.

There is also a practical argument that has nothing to do with law. Audiences increasingly spot AI video anyway, and being told is very different from finding out. The reputational risk of undisclosed synthetic media is larger than the reach cost of a label.

Consent covers whether you may use a likeness. Retention covers what happens to the material afterwards, and it is where enterprise procurement spends its time. Business buyers ask this first, which is why enterprise avatar tools are usually compared on retention terms rather than output quality.

Three questions carry most of the weight. How long are uploads kept after a generation completes. Are they used for anything other than producing your output. Can you get a written answer to both.

A vendor that keeps uploads indefinitely by default is not disqualifying on its own, but it should be a decision you made rather than one you discovered. If the answer is not in the privacy policy, it is not a commitment, it is a current practice that can change without telling you.

Only what we can point at, because a list like this is worthless if the vendor writing it exempts itself.

Do you require proof of consent for likeness and for voice separately. Is output watermarked, and can that be turned off. Was the model trained on licensed data. Can I delete a likeness, and does deletion propagate. Do you retain my uploads for training. If you would rather start from a shortlist, the generator comparison notes which products publish consent controls at all.

Vendors who answer all five clearly are your shortlist. Vendors who answer three and deflect on two have told you something as well.

If you are comparing platforms on capability as well as policy, our comparison of AI avatar generators covers the output side, and the pricing page covers what each tier includes.