An AI video file can be technically exportable and still legally or editorially unsuitable. Rights in source footage, photographs, brand material, music and performances are separate from a generator’s license to use its output. Synthetic visual evidence also introduces a problem of audience trust: realistic motion can falsely suggest that someone spoke, endorsed a product or appeared at an event.
Start from the inputs, not the output terms
Keep records for every source still, audio clip, video reference, script and recognizable person. Permission to show a photograph does not necessarily grant permission to train a reusable avatar, generate new facial performances or advertise a product. Permission to use a music track in one original video may not extend to derivative synthetic versions or international ad distribution. A local legal review should cover intended platform, country, language, duration and use.
Runway’s published usage-rights policy says creators may use generated outputs commercially without a mandatory Runway credit. Adobe’s guidance also permits commercial use under stated feature and beta conditions. Neither statement extinguishes outside claims from photographers, actors, licensors or trademark owners. Conversely, the Tencent HunyuanVideo community license has explicit excluded territories, reminding producers that a downloadable checkpoint is not universal permission to run it.
Address likeness and truthful representation
HeyGen and Synthesia document same-person live consent procedures for personal avatars. Those platform controls should be reinforced with a written purpose-specific release and a mechanism to stop future output. OpenAI’s historical Sora 2 system card discussed nonconsensual likeness and deceptive-media risks. A documentary must distinguish recreation and illustration from archived footage, especially in sensitive political, financial and medical contexts.
A consent ledger need not be public, but the production team should be able to trace a final clip to an approved source and a permitted use. An output that mimics an identifiable real person should never be presented as their actual recorded statement unless that characterization is true and authorized.
Capture provenance
Archive creation timestamp, source IDs, exact model and version, modifications and editorial reviewer. Where a provider supplies provenance or watermark technology, record its presence and limits. Google’s Veo guidance describes SynthID watermarking; that does not independently prove that every redistributed or transcoded copy retains easy-to-inspect provenance.
Publish with an audience in mind
A fictional animated scene may need little more than normal creative credits. A historical reconstruction, altered interview or synthetic spokesperson should be labeled when the synthetic nature changes how a reasonable viewer would interpret it. Avoid using video generation to fabricate testimony, evidence or advertising endorsements.
This is workflow education, not legal advice. AI 11 owns the in-depth policy and jurisdiction analysis; AI 04 documents where those concerns enter video generation, localization and edit approvals.
Worked production example
A synthetic recreation of a nineteenth-century market panic should be captioned as a reconstruction rather than filmed evidence. If it uses a specific historic person’s likeness, identify any relevant rights and portrayal concerns instead of presuming that age makes the image harmless. For a contemporary interview, substituting generated facial motion would be much more sensitive because it could alter perceived testimony. Retain a change log that shows where generative elements enter the edit, and ensure music or archival images have their own licenses regardless of the generator’s commercial-output terms.