Where AI is used, and who labels what
Effective date: 18 September 2026
AI questions: productions@lowlightking.com
AI is a margin lever with a named human gate. It is not the reason to hire us, and both local competitors already sell it as the headline. This page states where it is used, what is never generated, which party has to label what, and the one rule that a client's legal team asks about first: client material never trains a model.
How we use AI is the shorter, public version of this page. Where the two differ, this one governs, and clause 18 of the Terms governs over both in an engagement.
1) Where AI is used, and the gate on it
| Stage | What AI does | The human gate |
|---|---|---|
| Scripting | First draft, variant generation | The founder approves the narrative and every factual claim |
| Storyboard | Rapid concept boards | Approved before production |
| Voiceover | Synthetic voice for drafts and low-stakes internal assets | Human voice for all awareness, proof and close assets |
| Motion and explainer | Asset generation, background plates, transitions | The editor assembles; the founder signs off |
| Localisation | Multi-language variants | Native-speaker check before release |
| Derivatives | Cutdowns, captions, aspect-ratio variants | Spot-checked |
| Colour and finishing | None | Human, in DaVinci Resolve |
| Site imagery | The reference frames on this website are generated | Never a client, a client's premises, a real person, or a delivered asset. Every frame is logged with the prompt, the model and the date that made it |
The frames on this website are made with fal-ai/nano-banana-pro through
fal.ai, at build time. That tool never receives client material or personal
data: it is given a written prompt and returns an image, and the record of every
prompt is kept. It is not a processor and it is not in the subprocessor list,
because there is nothing of anyone's for it to process.
2) What is never generated
- A client's people. A face on screen belongs to a person who signed a release.
- A client's premises, offices or product, unless the client asks for a generated depiction in writing and it is labelled as one.
- A testimonial, a customer, a case-study outcome or a piece of view data.
- Anything a viewer would reasonably assume was filmed.
- A delivered asset in place of one that was shot, without saying so at handover.
These are not aspirations. Where a brief would require any of them, we say so before the shoot rather than at delivery.
3) Client material never trains a model
Client Materials, raw footage, working files, deliverables and the personal data we process for a client are never used to train, fine-tune, evaluate or improve a machine-learning model, whether ours or a third party's.
Where a production tool would send client material to a provider that trains on its inputs, that tool is not used on client material, or it is configured so that it does not, and we satisfy ourselves of that before the material reaches it. Clause 13 of the data processing agreement states the same obligation in a form a controller can enforce, and clause 18F of the Terms states it for material that is not personal data.
We do not sell, license or contribute client material to a training dataset, and no instruction from anyone at this company can authorise it.
4) Who labels what
Two regimes reach this work and they put the duty in different places. Neither party relies on the other to do what attaches to itself.
| Regime | What it requires | Who carries it |
|---|---|---|
| IT (Intermediary Guidelines and Digital Media Ethics Code) Amendment Rules 2026 — India, in force 20 February 2026 | Synthetically generated information must be labelled, carry permanent metadata or a provenance identifier, and be declared to the platform | Intermediaries and hosting platforms, reached through the uploader's declaration. In practice the uploader is the client |
| EU AI Act Article 50(2) — applicable 2 August 2026 | Generative output marked machine-readably and detectable as artificially generated | The provider of the generative system — ordinarily the tool vendor. We do not strip a mark a tool applies |
| EU AI Act Article 50(4) | Disclosure that deep-fake image, audio or video content has been artificially generated or manipulated | The deployer. Where we operate the tool, that is us. Where the client publishes the result, that is the client |
| EU AI Act Article 2(1)(c) | The Act follows the output, not the producer's address | A deliverable made in Kochi and distributed to buyers in the Union brings both parties into scope |
What we do on every project. Where a deliverable contains synthetically generated audio, video or audio-visual content, we tell the client in writing at handover which deliverables contain it and what was generated, and we embed a label or provenance metadata in the file, following the provenance standard the distribution platform accepts and naming which one we used. That written statement is what a client needs to make a platform declaration, and it is given whether or not it is asked for.
What is not synthetic content. Colour work, noise reduction, retouching, rotoscoping, upscaling, subtitling and translation do not make a deliverable synthetically generated, whether or not the tool doing them uses machine learning. Treating them as synthetic would make the label meaningless on the deliverables where it matters.
5) Likeness, voice and digital replica
A person's name, image, likeness and voice are protected in their own right, separately from copyright and from data protection law, and a growing number of jurisdictions protect them specifically against synthetic reproduction — Tennessee's ELVIS Act, California's AB 1836 and AB 2602, and Denmark's copyright amendment among them. A federal US statute has been proposed and is not law.
Our releases state the media, the territory and the term, and state expressly whether a likeness or voice may be used to generate, alter or train synthetic output. Where a release does not say so, no synthetic use is permitted, and no client instruction can authorise one the release does not. This applies to material a client supplies us as much as to material we capture. Clause 11.5 of the Terms is the contractual version.
6) AI literacy
Article 4 of the EU AI Act requires those operating AI systems to understand them well enough to use them properly. The people running these tools here are trained on what each tool does with an input, what it retains, whether it trains on what it is given, where its output can mislead, and which of the duties in Section 4 attaches to using it. A tool nobody has been through that with is not used on client work.
7) Asking what was used
Write to productions@lowlightking.com with “AI” in the subject line and name the deliverable. We will tell you which stages used a tool, which tool, and what a human changed afterwards. We acknowledge within 24 hours and answer within 30 days, the same commitment the grievance mechanism in the privacy notice makes.
We do not hold that information back as confidential. A client who cannot find out how their own asset was made cannot answer their own regulator.
8) What we do not claim
We hold no certification or conformity assessment under the EU AI Act, and none is required for the transparency duties in Section 4 — they are not high-risk obligations and no listed high-risk use case applies to this work. We do not offer legal advice on a client's own AI obligations, and a client distributing into a regulated sector or a jurisdiction we have not named should take its own.
The other documents
This is one of nine. The index lists them all and says what each one decides.