AI-generated video can look striking, but commercial use exposes limits that a demo reel hides: shots that do not match each other, garbled text and logos, implausible physics, unsettled copyright, and platform rules about labelling. Most of these can be managed with the right workflow, and some cannot. Knowing which is which lets you decide where AI belongs in your campaign and where it does not.
This article separates technical limits from legal and platform limits, then lists mitigations. It is general information, not legal advice.
Technical limitations
These are common characteristics of current generative video, not permanent facts. Tools improve quickly, so test any claim against the tool your studio uses today.
Consistency across shots
A commercial needs the same product, character, wardrobe and setting from shot to shot. Models generate each clip separately, so details drift: a logo moves, a face changes, a jacket gains a pocket. Some tools now accept reference images to help. Google’s documentation, at the time of writing (November 2026), describes character consistency across up to six reference images for one model (Google AI Studio documentation), which helps but does not guarantee identical results for every shot.
Mitigation: use real references, keep shots short, keep a shot bible, select takes by accuracy, and cut around problems.
Hands, text and logos
Fine structures such as fingers, small type and exact logo geometry are common trouble spots. A label that reads as nearly right is worse than none, because audiences and regulators notice.
Mitigation: never rely on generated text. Add logos, packaging text, prices and legal lines as real graphics in the edit. Frame hands carefully or film the real action.
Physics and motion
Liquids, fabric, reflections, object contact and complex movement can behave oddly. Objects may pass through each other or change as they move.
Mitigation: favour simple actions, check every frame at full size, and film or simulate the moments where physical believability is the point.
Fine product detail
Textures, small components and precise proportions can shift. That matters when the product is the star. See our guide to AI product videos for a pipeline built around real packshots.
Duration and complexity
Clips are short, so long scenes are assembled from many generations. Anything with continuous complex action or dialogue over many seconds is harder.
Mitigation: design the edit around short shots, and plan sound separately.
Control and revision
You steer with prompts, references and selection, which is less precise than directing a person on set. Small changes may require regenerating a shot, which can alter the parts you liked.
Mitigation: lock the storyboard first, approve in stages, and keep generated and real layers separable.
Legal limitations
Copyright in the output
Ownership of AI-generated material is an open question in many jurisdictions. In the United States, the Copyright Office concluded that AI outputs are protectable only where a human author contributed sufficient expressive elements, that prompts alone do not make someone the author, and that AI-assisted works are not barred from copyright (US Copyright Office). That is a US position.
In Australia, the Attorney-General’s Department sought feedback from its Copyright and AI Reference Group between October and December 2025 on priorities that include greater certainty about how copyright law applies to material generated with AI (IP Australia). The sources opened for this article do not show an Australian statute or court decision that resolves it, so treat the position on purely generated footage as unsettled and get advice.
Practical effect for a brand: the more human creative contribution in a commercial (script, selection, edit, graphics, sound, live action), the stronger your position is likely to be. Purely generated clips may be harder to protect against copying. Contract terms with your studio matter. See international content licensing for licence basics.
Training data
The Australian Government said in October 2025 that it would not provide a text and data mining exception to copyright for AI developers, and the reference group has been looking at licensing arrangements and clearer rules for AI tools (iTnews). For brands, the unresolved question is how models were trained and what that means for outputs. Ask your studio which tools it uses and what the provider’s commercial terms say, including any indemnities.
Likeness, voices and music
Do not generate a real person’s likeness or voice without clear permission. Do not assume generated music is cleared for your use. Get written licences.
Australian Consumer Law
The ACCC says that images and descriptions must be accurate and based on reasonable grounds, applies to advertising and social media, and that intent does not matter (ACCC). A generated product that looks better than the real one, or a generated result presented as real, creates risk.
Platform AI-label policies
Policies change, so check each platform before you publish.
- YouTube asks creators to disclose realistic altered or synthetic content, such as realistic scenes that did not occur. It does not require disclosure for non-realistic content such as fantastical or animated visuals, or for minor aesthetic edits (YouTube Help).
- Meta may require a label for photorealistic video or realistic-sounding audio created or altered digitally, including with AI, and says there may be penalties if a required label is missing. It may also apply an “AI info” label when its systems detect industry-standard signals (Meta Help).
- TikTok has said it automatically labels AI-generated content that carries C2PA Content Credentials metadata (TikTok Newsroom). This article does not summarise TikTok’s creator labelling rules, so confirm its current requirements for creators and advertisers directly.
C2PA is a coalition-created technical framework, branded Content Credentials, that records where content came from and how it was modified (C2PA). Because labelling can be triggered by metadata, keep a record of what was generated and avoid stripping provenance data from files.
A risk register you can use
| Risk | Likelihood to check | Mitigation | Owner |
|---|---|---|---|
| Product looks different from real item | High in product films | Real packshot references, side-by-side review | Producer |
| Garbled text or logo | High | Real graphics layer | Editor |
| Shot-to-shot drift | High | References, shot bible, selection | Creative lead |
| Physics glitches | Medium | Simple actions, frame-by-frame check | Editor |
| Copyright position unclear | Ongoing | Human contribution, contract terms, advice | Client and studio |
| Missing platform label | Medium | Per-platform disclosure plan | Campaign manager |
| Misleading claim | Medium | Claims file, ACL review | Client |
When AI is the wrong choice
Choose a conventional shoot, or a hybrid, when the film depends on exact product appearance, a real person’s authentic testimonial, regulated claims, or sustained complex action. See AI commercial production vs traditional filming for a decision table. For the broader picture, start with AI video production for brands.
Where Radiance Impact fits
Radiance Impact is a Brisbane-based creative technology studio. Its Studio produces AI commercials and product films, brand films and trailers, post-production, and voice and sound. If you want an honest view of whether AI suits your project, contact us or see the Studio page.
Sources
- Gemini API video generation documentation, Google AI Studio
- Copyright Office Issues Part 2 of Artificial Intelligence Report, US Copyright Office
- Key copyright policy priorities, IP Australia
- Government rules out AI training exemption in copyright laws, iTnews
- False or misleading claims, ACCC
- Disclosing use of altered or synthetic content, YouTube Help
- How to identify AI content on Meta products, Meta
- Partnering with our industry to advance AI transparency and literacy, TikTok Newsroom
- Coalition for Content Provenance and Authenticity, C2PA
