AI Disclosure

What This Page Tells You

Every video Clipwright returns is generated. The person is not a person, the voice is not a recording, and the lip movement is computed. Nothing we produce is footage of anything that happened.

Two things follow, and they are not the same thing: we mark the file so that a machine can tell, and you may be required to mark the publication so that a viewer can tell. The second obligation is yours, and this page explains why.

What Is Generated and by Whom

• The voice is synthesised from your script by ElevenLabs. It is not a recording of anyone reading it
• The presenter is produced by HeyGen — either a synthetic actor or an animation driven by the image you supplied
• The lip sync is computed to match the synthesised audio
• The script is yours. We do not write it, and we do not check it

The Privacy Policy names these vendors as subprocessors and says what reaches each of them.

The presenter can also be an actor you asked us to make. Where you request one, an image model at fal.ai draws a portrait from a description you wrote in words, and that portrait drives the presenter in every video you then make with it. No photograph is involved and no real person is depicted: describing one is prohibited by the Acceptable Use Policy, and we screen the description before anything is drawn.

How We Mark the Output

Every finished video is marked machine-readably as generated. The statement is fixed, and it reads:

This video was generated with AI: the actor, the voice and the lip sync are synthetic.

It travels two ways, deliberately. The object in our storage carries the statement as storage metadata alongside ai-generated: true and generator: clipwright, and the file itself carries the mark embedded in its own metadata. The API returns the same fields with every finished run.

Why two. The embedded mark survives download but can be stripped by re-encoding; the storage metadata is applied every time and cannot be lost by our own pipeline, but does not follow the file out. Neither is sufficient alone, and we do not claim either is tamper-proof — a mark inside a file is a statement, not a lock.

The embedded mark depends on how the file is laid out, and we say so when it could not be written. Some encodings leave no place for it. The run still succeeds and the video is still returned — with a warning on the run saying the tag was skipped, so you learn it from the API rather than by inspecting the file. The storage metadata and this page are unaffected: the clip is generated either way, and that is what you are required to disclose.

Your Obligation When You Publish

Article 50(4) of the EU AI Act addresses the deployer — the person who publishes the content — and not the tool that produced it. If you publish a deep fake, meaning content depicting a person appearing to say or do something, you are required to disclose that it is artificially generated, in a manner that is clear and distinguishable at the moment the content is presented.

A machine-readable mark inside the file does not satisfy that. Viewers do not read file metadata. Platforms differ in what they do with it, and several strip it on upload. What satisfies the obligation is a label the audience can see.

Most platforms now provide their own control for this — a “synthetic media” or “altered content” switch at upload. Using it is usually the simplest route. Whether you use theirs or ours, the duty is yours.

The Visible Label We Can Add

Because a visible label is what the obligation actually needs, the API can burn one into the video. Pass disclosure_overlay: true with a run and the finished clip carries the statement on the picture itself, where re-encoding and re-uploading cannot remove it.

It is off by default. We do not add it silently, because a caption burned into your video changes what you receive, and the request decides what you receive.

It depends on the composition stage, and a deployment can have that stage switched off. Where it is off, a run asking for the overlay is refused before anything is charged — the API answers that the overlay is unavailable and says the output is still marked in the run contract and in the file metadata. We say this here rather than promising the feature flatly, because a page that claims what a deployment may not offer is worse than one that names the condition.

What We Do Not Claim

• We do not watermark cryptographically. There is no C2PA manifest and no signature you could verify against us. The mark is a declaration, and a determined person can remove it
• We do not detect misuse. We do not review scripts before rendering and do not scan the output. What we prohibit is set out in the Acceptable Use Policy, and it is enforced on report, not by inspection
• We do not track where you publish. We cannot tell whether you labelled the publication, and this page is not a claim that you did

Contact Information

Clipwright is operated by Dimantika sp. z o.o., Gdansk, Poland. Questions about this page, or about a video you believe was published without the disclosure it needed, go to support@dimantika.com.