AI policy. How I use AI in video, and what I disclose
The short version. I label generated or materially altered content whenever a reasonable viewer could take it as real. I never synthesise a real person without written consent specific to that use. AI that only helps the humans, transcription, denoising, colour and edit tools, gets no label. Two questions decide everything else, and they are below.
I make films the traditional way, with crews, lenses and real light, and I also make work that is partly or entirely AI-generated. Because I work on both sides of that line, I get asked the same two questions on almost every job. Do we have to tell the audience, and what do we need from the people in it?
This page is my working answer. It is not legal advice and it is not another industry framework. There are good frameworks now, and I point to them below. What has been missing is the practical layer, what a filmmaker or a commissioning team actually does on a live production. This is the policy I apply to my own work and the one I recommend to clients. It is written to be usable by a team without a legal department, and I review it as the rules change.
The two questions that decide most cases
Nearly every real situation resolves through two tests.
The contributor test. Could I show the person exactly how their likeness, voice or performance has been used, and would they be comfortable with it? If I could not show them, or I suspect they would not be comfortable, the use needs consent I do not yet have.
The audience test. Does the AI change the audience's understanding of what they are looking at? If a viewer would read something as real that was generated, or believe a person said something they did not, the audience needs to be told, clearly and where they will see it. Where that is depends on the room they are watching in, which is the subject of where video lives now.
If a use passes both tests, it almost never needs more process. If it fails either one, it always does. The rest of this page is what that process looks like.
The case that fails both tests
Here is the shape of the failure I see most often. A company needs a customer testimonial. A real customer records her lines in a voice session, genuinely enthusiastic. The agency takes a few stills and the recording, feeds them into a generative system with scripted answers, and produces a lifelike sixty-second piece. The eyes blink, the mouth syncs, the light catches her glasses when she leans in. It looks like she sat in a studio and answered live. She did not.
It fails the contributor test, because she was never shown what would be made from her. It fails the audience test, because the piece only works if the viewer believes it was filmed. Months later someone puts a side-by-side online, and the story stops being the product and becomes whether the brand tried to pass something off. A version that holds up can still use AI. It just stays honest about what was captured and what was constructed.
Three tiers of AI use
I classify AI use on any production into three tiers. This lines up with how the industry is starting to label work, including the three-tier approach in the Human Provenance in Film standard launched at Cannes in 2026, so nothing here fights the direction the industry is moving in.
| Tier | What it covers | What I do |
|---|---|---|
| Assisted | Transcription, noise reduction, colour tools, edit assistance, research and drafting. AI helped the humans and no generated content appears in the finished work | Covered by normal production practice. No label, no special consent. Tool choices still follow the data rules below |
| Altered | AI materially changes something real. A reconstructed shot, an extended frame, a cleaned-up or re-voiced line, a changed appearance | Specific agreement from anyone affected, and a label wherever the audience test says the change could mislead |
| Synthetic | AI generates content presented in the finished work. Generated scenes, synthetic voices, digital doubles, AI presenters | Explicit written consent from anyone represented, and clear disclosure to the audience, on screen where it matters |
The tier is decided by what ends up in front of the audience, not by which tools were open during the edit. Denoising an interview is assisted. Generating an establishing shot is synthetic, even if it took five minutes.
What I disclose, and how
My standing rule is the audience test applied plainly. I label generated or materially altered content when a reasonable viewer could otherwise take it as real. I do not plaster a warning on every piece of work because an AI tool touched the timeline somewhere, because that empties the label of meaning.
In practice that means
A visible on-screen note on or beside any synthetic shot that could read as real. The wording I use is "AI-generated imagery" or "AI-assisted reconstruction", short enough to read, placed where the shot appears rather than buried in credits.
A standing statement on projects that mix real and generated material, at the head or foot of the piece. My usual wording is "This film contains AI-generated imagery where indicated. Everything presented as real footage is real footage."
No label for assisted work. Captions, denoising, colour, edit tools.
Never using disclosure as a cure for something misleading. A label does not make a fabricated endorsement acceptable. If the piece only works because the audience is deceived, the piece does not get made.
If you want the whole section as a rule for a brief, it is four lines.
If it is proof, capture something real. If it is synthetic, label it where people will actually see it. If it uses a real person, get consent that includes synthetic use. If it is scripted, say it is scripted.
What I ask of contributors
The principle first, because it settles most of it. Permission to film someone is not permission to synthesise them. A signed release for a shoot does not license a voice clone, a digital double, or new footage of that person doing things they never did. Those need their own consent, in writing, specific to the use.
So on any production where the altered or synthetic tiers touch a real person, the consent covers
Exactly what is being used. Face, voice, performance, recordings, photographs.
Exactly what will be made with it, and where it will appear.
How long the permission lasts, and that materially different future uses need fresh consent.
Whether the material may ever be used to train or improve an AI model. My default is that it may not, unless expressly agreed.
How consent is withdrawn, and honestly, what can and cannot be unpublished once work has shipped.
I keep a consent record per production. It holds the original source files, the signed releases including any synthetic-use consent, the exact disclosure wording used and where it appeared, a short log of which AI tools were used and for what, and the approvals, even if that is an email thread. If a question comes up later, the answer is written down, not remembered.
If you need a synthetic-use add-on for a release, keep it narrow and specific. This is the shape I use. It is a starting point, not legal advice.
The participant consents to their voice and likeness being used to create AI-assisted or AI-generated video and audio that may include synthetic performance and a generated setting. This consent applies only to the specific project, versions and channels agreed in writing. No further reuse, no training of any model, and no creation of new performances using the participant's voice or likeness is permitted without further written consent.
For anyone under 18 or in a vulnerable position, the bar goes up. Guardian approval where the law requires it, plain explanation, and no reliance on the pressure of a casting or workplace relationship.
Tools, data and provenance
Two quiet rules that prevent most problems.
Nothing identifiable goes into a consumer AI tool without checking its terms. Rushes, voice recordings, scripts and client material do not get uploaded to services that retain them or train on them. I keep a short list of tools whose terms I have actually read, and that list travels with the production.
Provenance gets preserved where the tools support it. Content Credentials (C2PA) and watermarking such as SynthID are becoming the plumbing of trust in video, and the EU now requires providers of generative systems to mark output in a machine-readable way. Where a tool can attach provenance, I leave it attached.
Who owns generated output is a separate question with a different answer, and I deal with it in who owns the IP, and how do you prove you made it.
Where this sits with the law
The short version, current as of September 2026, and the reason this page carries a review note rather than a publish-and-forget date.
In the EU, Article 50 of the AI Act has applied since 2 August 2026. Anyone deploying AI-generated or manipulated image, audio or video that could falsely appear authentic must disclose it clearly. It was not delayed with the high-risk provisions. Fines for breaching it run up to 15 million euros or 3 percent of worldwide turnover. The one runway is for the providers of generative tools, who have until 2 December 2026 to mark the output of systems already on the market in a machine-readable way. If your audience includes the EU, this is a present obligation, not a future one.
In the UK, there is no single law that requires an AI label. The ASA's position is that the advertising codes are media-neutral, so an ad that would mislead without a label breaks the rules however it was made, and a label does not cure a misleading claim. Ofcom has set out transparency measures for synthetic media, labels, provenance metadata and context notices.
On the data side, the ICO's tests for consent apply. Freely given, specific, informed, unambiguous. Processing a face or voice for identification can trigger the stricter biometric rules, which usually means explicit consent and an impact assessment.
Industry frameworks are converging on the same shape. The IAB's disclosure framework for advertising is materiality-based, label when AI affects authenticity or identity, not on every use of a tool. Human Provenance in Film offers a three-tier label for finished films. BAFTA has updated its rulebook to reward human creative achievement regardless of the tools used, and reserves the right to ask about the nature of any AI use and the extent of human involvement. My policy is deliberately compatible with all of them. It answers the question they leave open, which is what a working production does on the day.
Questions
Do I have to disclose AI-generated video in the UK? There is no single UK law that says label it. What applies is the ASA's position that the advertising codes are media-neutral, so an ad that would mislead without a label breaks the rules whatever made it, and the ICO's rules on consent wherever a real person's data is involved. If any of your audience is in the EU, Article 50 of the AI Act has applied since 2 August 2026 and requires clear disclosure of AI-generated or manipulated content that could pass as real. My rule covers all three. Label whenever a reasonable viewer would otherwise take it as real.
Does a filming release cover a voice clone or a digital double? No. A standard release licenses the footage you shot and the ways you agreed to use it. It does not license a cloned voice, a synthetic likeness or new performances the person never gave. Those need their own consent, in writing, specific to the use, with a stated duration and a stated position on training. Sample wording is above.
What should a consent form for AI use of someone's likeness cover? Five things. What is being used, face, voice, performance, recordings or photographs. What will be made with it and where it will appear. How long the permission lasts, and that new uses need fresh consent. Whether the material may train or improve a model, with no as the default. How consent is withdrawn, and honestly, what can and cannot be unpublished once the work has shipped.
Do you make fully AI-generated video? Yes, where it is the right tool for the piece and the audience is not misled about what they are watching. I also regularly advise clients that AI is the wrong tool for their piece. Sometimes the right answer is a crew and real light, and I will say so.
Does using AI make video cheaper? Sometimes, for the right kind of content. It does not make judgement cheaper, and judgement is usually what was being paid for. The expensive failures I see are not tool failures. They are pieces that lost the audience's trust to save a shoot day.
Do I have to label a video just because AI helped edit it? No, and I would advise against it. Labels are for uses that change what the audience believes they are seeing. Labelling everything trains audiences to ignore the label that matters.
Can you help us write our own policy? Yes. Adapting this into a policy, consent forms and a commissioning checklist for a team is part of the advisory work I do. The checklist starts with five questions. What source material was used, and is it licensed or consented. Where will disclosure appear, and will viewers actually see it. Do we have explicit consent for likeness and voice, including synthetic performance. What are we keeping on file if someone asks how it was made. Can we produce a real alternative for anything that functions as proof.