How to use AI ethically for video generation. A working standard for filmmakers, producers and brands

AI Video Tools
Written By Nigel Camp

Updated 18 September 2026

Using AI ethically in video comes down to one question. Has the tool changed who appears to speak, what they appear to say, or what appears to have happened. If it has, you need consent from anyone shown and a label the audience will actually see. If it hasn't, most AI use is workflow, and workflow doesn't need a disclosure.

Picture a situation. Sally works in marketing for a mid-sized company. The team is building out a product campaign, budgets are tight, and someone suggests cloning her voice rather than booking a session every time a script changes. Sally says yes. She's flattered, honestly, and nobody in the room is acting in bad faith. There's an email thread, a quick approval, and the campaign ships.

A year later Sally leaves for another job. The voice stays. It's still narrating product videos for a company she no longer works for, in a tone she never chose, saying things she was never asked about. And then she hears something that sounds a lot like her reading a script for a different brand entirely, and nobody can quite explain how that happened or who at the vendor holds the model.

Nothing in that story required a villain. It needed one thing that wasn't there, which was a written agreement naming what the voice could be used for, by whom, and for how long. The UK has no statutory right over your own likeness, so the only protection Sally ever had was the one the contract gave her. There wasn't a contract, so there wasn't any.

That's the shape of most AI ethics failures I see coming. Not deepfake fraud or malicious impersonation, but ordinary teams under deadline pressure making a reasonable-sounding decision that nobody wrote down properly.

I've made films for brands and charities for fifteen years, most of it filmed the traditional way. Most of what I'm making right now is still human-led storytelling, with AI turning up in a graphic or a clean-up rather than in the story itself. But organisations are leaning further into the creative side of it every quarter, and I see those decisions from the inside rather than the outside. That's the vantage point this page is written from. Not a vendor with an avatar product to sell, and not a commentator who's never had to get a cut approved.

It's also the working version of my published AI policy. What the rules now require, when a label protects trust and when it just creates noise, and the lines I won't cross regardless of what a tool makes possible.

Explore this guide

If there's one section to start with, begin with When disclosure matters and when it just creates noise. The four tests there decide everything else on this page.

Where I stand

Most AI in a professional timeline doesn't need a label. Transcription, noise reduction, colour matching, captions and clean-up don't change what anyone said or did, and labelling them tells the audience nothing. Plastering a disclosure on every caption job empties the word for the moments it matters.

A label starts to matter when AI changes the substance of what a viewer believes. A generated presenter, a cloned voice, a reconstructed scene, an altered testimonial. That's where you disclose, and where consent stops being a courtesy and becomes the foundation of the job.

There's a quieter version of this that has nothing to do with technology, and it's worth naming because AI makes it easy at scale. If someone on screen is telling you how much they love a product, they should be someone who has actually used it. A generated face delivering a real customer's words, an actor's performance stitched onto a genuine review, a composite testimonial assembled from three people's feedback. Each of those is a small step, and each one moves the work from documenting something into manufacturing it. The disclosure question and the authenticity question aren't the same question, and passing one doesn't get you through the other.

The same goes for how organisations treat their own people. Staff aren't talent with an agent and a rate card. When the person being asked to hand over their face or voice depends on the person asking for their next appraisal, "they said yes" isn't a high enough bar. Get it in writing, name the uses, put an end date on it, and tell them what happens to the model when they leave.

Assisted, altered or synthetic. Where your AI use sits

AI can speed up editing, draft scripts, clean up audio, generate subtitles, and produce synthetic voices and faces. For a lot of teams that means faster, cheaper content across social clips, product explainers, training films and internal comms. The catch is that those uses sit at wildly different ethical altitudes, and treating them as one thing called "using AI" is how teams end up disclosing the wrong things and missing the ones that matter.

So before you decide whether to disclose anything, work out what kind of AI use you're actually looking at. I sort every job into three tiers, and they line up with the provenance tiers now emerging in the film industry through Human Provenance in Film, launched at Cannes in May 2026 and still in consultation until 31 October 2026.

My tier What it covers HPF equivalent What the viewer sees
Assisted Transcription, clean-up, colour work, translation, edit suggestions. The tool touches the workflow, not the meaning Assistive AI Nothing. No label needed
Altered Generative removal or replacement, voice modification, scene extension. Supplied material, materially changed Usually Assistive AI. Generative AI if it originates picture, voice or performance A label where the change could mislead
Synthetic Generated people, voices, places or events. Material that never happened in front of a camera Generative AI An on-screen label where it matters, and written consent from anyone represented

Traditionally captured footage sits underneath all three as the baseline, and the usual questions apply to it. Was it lawfully recorded, and do you have the rights.

One caution about the tiers. They're not moral labels. Assisted isn't automatically safe, because a "clean-up" that removes a protester from a wide shot has changed what appears to have happened. And synthetic isn't automatically wrong, because a labelled, consented reconstruction can be the most honest way to show something no camera caught. The tier tells you which questions to ask, not what the answer is.

The minimum standard before anything goes live

Ethics can sound abstract until you try to ship a video under deadline pressure. So here's a baseline that works for brand video, internal comms, recruitment, training and social content. It's intentionally practical, and it sits comfortably alongside NIST's AI risk management framework for anyone who needs a formal organisational reference behind it. If you meet this minimum, you reduce risk without strangling the creative process.

  • Confirm consent and rights for any identifiable face, voice, or likeness

  • Disclose meaningful synthetic changes that could mislead a reasonable viewer

  • Keep a basic log of what tools were used and what was altered

  • Run one independent review for bias, tone, and reputational risk

  • Verify provenance for any third-party clip before you reuse it

This isn't about being perfect. It's about being accountable. If you can't explain what changed and why, that's usually a signal the workflow needs tightening.

When disclosure matters and when it just creates noise

This is the bit most ethical AI guides skip. They say be transparent but never define what that means. A useful rule is to separate workflow assistance from meaning and identity. Disclosure usually isn't necessary when AI is doing background work that doesn't change the substance of what the viewer sees or hears. Think subtitles, light noise reduction, or removing a hum.

Disclosure becomes important when AI changes any of the following.

The four tests

Has AI changed what the viewer would believe

Run any AI use through these four. One yes is enough to mean it needs a label, and usually consent as well.

  1. 01

    Who appears to be speaking

    A generated presenter or a cloned voice changes the answer. Disclose it, and get written permission from anyone whose voice or face is used.

  2. 02

    What someone appears to have said

    Reordering or synthesising words puts a statement in someone's mouth. If they would not recognise the version you have built, it is not theirs.

  3. 03

    What appears to have happened

    Generated or altered events that read as a record of a real place or moment need a label, wherever the audience could take them as coverage.

  4. 04

    Whether the person shown is real, present or consenting

    The highest risk of the four. Without consent that names synthetic use, in writing, the shot does not go in.

No to all four. You are looking at workflow, not meaning. Transcription, captions, noise reduction, colour work and clean-up do not need announcing, and labelling them empties the word for when it matters.

If the viewer could reasonably believe something is authentic when it isn't, that's the moment where a simple disclosure protects trust. I call these the four tests, and every scenario on this page runs through them. The thread linking all of it is trust. Once a viewer starts wondering what's real, the content stops doing its job, whatever else it had going for it, which is why trust and authorship now need rules people can feel.

The UK's advertising regulator lands in the same place. CAP's guidance on disclosing AI in ads, published in May 2025, says there's no blanket duty to disclose AI in UK ads, that the codes apply however an ad was made, and that a disclosure won't cure an ad that's misleading. In other words, the question was never "did you use AI". It's "would the audience be misled if you stayed silent", and a label on a claim the pictures can't support fixes nothing. That's the four tests in regulator language, and it's worth knowing the two match.

Where the label actually lives

Most guides treat disclosure as a yes or no. On a real job the harder question is where the label goes, because there are five surfaces and they do different work.

  • On screen, at the moment of the synthetic shot. The strongest form, and the one I default to when a generated element could be mistaken for the real thing

  • In the description, end board or credits. The documentary convention, and where AI tools get listed the way archive and music always have been

  • In the platform's own toggle. YouTube, TikTok and Meta each have one, and the rules table further down covers what each requires

  • In the file itself, through Content Credentials. Worth attaching where your tools support it, though don't rely on it, because a normal export or transcode often strips it

  • In the delivery paperwork. Anything heading toward a fund, festival, insurer or buyer increasingly travels with a declaration of what was generated

For the wording itself, I keep it short and specific. "Includes AI-generated imagery" against a synthetic sequence, and "AI-assisted reconstruction" where real events were rebuilt with generative tools.

The calls you'll actually have to make

Roadrunner and the big deepfake frauds get the headlines, but they're not what lands on your desk. These are, and all five run through the four tests.

A gap in the B-roll, filled with generated footage. If the shot is illustrative and nobody could mistake it for the event you're documenting, that's ordinary picture-making and it needs no label. If it's cut in among real footage of a real place so it reads as coverage of that place, you've changed what appears to have happened. Label it or shoot it.

A stumble or a heavy accent smoothed out of an interview. Removing an "um" is editing and always has been. Reshaping how someone speaks is different, because accent carries class, region and origin, and smoothing it is a decision about how that person gets to be seen. Ask them first. Most people say yes and are glad to have been asked.

A passer-by removed from a shot. Usually fine, and we've done it with rotoscoping for thirty years. It stops being fine the moment the thing you remove changes the meaning of the frame, and the classic example is tidying a crowd until an event looks better attended or calmer than it was.

A testimonial tightened up. Cutting for length is normal. Reordering sentences so a hesitant endorsement becomes an enthusiastic one is putting words in someone's mouth, and generative tools make that seamless in a way a jump cut never was. If the person wouldn't recognise the version you've built, it isn't theirs any more.

A founder's voice cloned so scripts can change without another session. This is the Sally problem and the highest-risk item on the list. It's doable, it's sometimes the right call, and it needs written permission naming the uses, a defined end date, and a plan for what happens to the model when they leave.

How one of those plays out start to finish

Picture a situation. A charity needs a two-minute film about a night shelter for a winter appeal. The budget covers one shoot day, and access to the shelter itself is restricted for obvious reasons, so there's no chance of filming residents at close quarters.

AI could solve that in an afternoon. Generate the interior, generate the faces, generate a voice reading testimony gathered by the caseworkers. It would look good, and the appeal would run on time.

Here's where the four tests bite. Generated residents change who appears to be speaking and whether the people shown are real, present or consenting. Testimony in a synthetic voice changes what someone appears to have said. That's three of the four, on the most sensitive subject a charity has, in a film whose entire purpose is to be believed.

So what actually happens is this. Interviews with two caseworkers and one former resident who gives informed consent, filmed properly. Anonymity handled in camera with framing and lighting rather than generated faces. A short sequence of the building's exterior and the street at night, real. Where the film needs to show what a shelter night feels like without filming residents, it uses a deliberately stylised graphic sequence, obviously illustrative, labelled on screen as an AI-assisted reconstruction. AI does the transcription, the captions and the grade assist, and none of that gets mentioned because none of it changes anything.

What goes in the folder is the consent form naming the reconstruction, the log of which tools touched which shots, a note of the label wording and where it appeared, and the source files for everything presented as real. That's about twenty minutes of admin, and it's the difference between being able to answer a question in two years and hoping nobody asks.

What the rules require right now

Rules checked 17 September 2026. Platform policies and regulation are moving quickly, so follow the source before you rely on any of this for a live job.

Who What it says What it means for your edit
EU AI Act, Article 50 Applies from 2 August 2026. Providers mark synthetic output machine-readably, deployers disclose where the rules require. The artistic exception is read strictly, and the advertising examples in the Commission's final guidelines of 20 July 2026 all need the standard label Work seen in the EU may be in scope without you targeting the EU. The duty follows decision-making control, so settle it in the contract. Fines up to €15m or 3% of worldwide turnover
EU AI Act, timing No retroactive labelling for video, image or audio generated before 2 August 2026. The cut-off is the date of generation. Legacy generative systems have until 2 December 2026 for machine-readable marking Material generated before August 2026 doesn't need retro-labelling. The publication-date rule covers text on public-interest matters, not video
CAP and the ASA, UK Guidance published May 2025. No blanket duty to disclose AI in UK ads. The codes apply however an ad was made. A disclosure won't cure an ad that's misleading Ask whether silence would mislead, and whether a label clarifies the claim or contradicts it
YouTube Creators disclose realistic altered or generated content. From 27 May 2026 the label sits under the player on long-form and as an overlay on Shorts, and YouTube auto-applies it on significant photorealistic AI where a creator hasn't disclosed. Animated, stylised and minor edits stay in the description Disclose in the upload flow, and assume the label may be applied for you if you don't
TikTok Creator label for realistic AI required since September 2023. C2PA auto-labelling since 9 May 2024, the first platform to do it. An auto-applied label can't be removed. The advertising policy is separate and requires disclosure on realistic AI ad creative Organic and paid are governed differently. An undisclosed AI ad can be rejected or restricted
Meta A disclosure tool is required for organic photorealistic video or realistic audio that has been digitally created or altered. Announced February 2024, rolled out April 2024, across Facebook, Instagram and Threads. Meta also reads C2PA and IPTC provenance signals Use the tool at upload, and don't rely on provenance surviving your export
UK, still pending The government's Report on Copyright and Artificial Intelligence of 18 March 2026 promised a consultation on digital replicas and a labelling taskforce with an interim report due autumn 2026. Neither had published as of 17 September 2026 Nothing to comply with yet. The UK has no statutory likeness right, so consent is contractual and the contract has to name synthetic use

The practical version is shorter than the table. Disclose in the platform's own tool at upload, because the platforms increasingly apply a label whether you disclose or not, and a label you chose reads better than one you were given. Assume anything seen in the EU may be in scope even if you didn't target the EU. Put the labelling duty in the contract, because under Article 50 it follows whoever holds decision-making control over the AI use, and on a commissioned job that's a question worth settling before the shoot rather than after the edit. And remember that a platform's rules can be stricter than the law, which means compliance is the floor and not the ceiling.

What the industry guidelines say

For documentary, the reference point is the Archival Producers Alliance guidelines, published in September 2024 after consultation with hundreds of producers. The core of it is that generative material should never be passed off as primary source, that AI elements belong on a cue sheet the way archive and music always have, and that AI use should be listed in the credits. It's documentary-specific and US-led, but the cue sheet idea travels to any kind of commissioned work.

In the UK, the anchor is the BFI and CoSTAR Foresight Lab report on generative AI in the screen sector, published in June 2025. Its sixth recommendation is public transparency, on the basis that transparency is what drives audience trust, and it cites a YouGov survey in which 86% of British respondents supported clear disclosure when AI is used in media production. That figure is worth sitting with, because it's a long way from the position that audiences don't care.

On the performer side, SAG-AFTRA's framing is the cleanest three words available, which are consent, compensation and control. It's US and contractual rather than statutory. In the UK, Equity is campaigning for personality rights covering likeness, voice and movement, and won AI protections in the TAC-Equity agreement for independent television production in Wales in June 2026. Until something statutory exists here, the contract is the only protection a performer or a member of staff has, which is worth remembering when you're the one drafting it. I've written separately about consent and likeness in AI film production, and that's where the release wording and the digital-double questions live.

How to tell whether an AI video tool is ethically sourced

The honest answer to "which AI video tool is the most ethical" is that a tool isn't ethical or unethical on its own. The use is. But tools do differ in how they were built, and if that matters to you or your client, there are four questions that separate them, and they're stable even though the product names change monthly.

Ask what the model was trained on, and whether the vendor will say so in writing. Some now advertise licensed or owned training data, and that claim belongs in your paperwork if it's part of why you chose them. Ask whether output carries Content Credentials, so provenance travels with the file at least until an export strips it. Ask what happens to the material you upload, because footage of your client's staff shouldn't quietly become training data. And ask whether the tool lets you meet your own obligations, meaning machine-readable marking under Article 50 and a workable route to the platform toggles. A vendor who can't answer those four in plain English is telling you something useful.

Three times the audience found out

Everything above is easier to hold onto when you see what happens without it.

In Roadrunner, the 2021 documentary about Anthony Bourdain, the director used an AI clone of Bourdain's voice to read words Bourdain had written but never spoken aloud. The words were real. The voice wasn't, the audience wasn't told, and reporting at the time made clear how contested the consent was. The film's craft didn't save it from the story becoming the story.

In What Jennifer Did, a 2024 Netflix true-crime documentary, journalists spotted that photos presented as archival showed the hallmarks of AI manipulation. Nothing was disclosed. In a genre whose only currency is that the material is real, generated "archive" isn't a shortcut, it's a hole in the hull, and it's a large part of why the APA guidelines exist.

And in January 2024, an employee at the UK engineering firm Arup transferred around 25 million dollars after a video call with what appeared to be the company's CFO and colleagues. Every person on that call was a deepfake, as the Financial Times later confirmed. That one isn't a disclosure story. It's what the same technology does in the hands of people who were never going to disclose, and it's why provenance habits belong inside companies and not just on film sets.

There's a counter-example worth one paragraph. When Steven Soderbergh premiered John Lennon, The Last Interview at Cannes in May 2026, he said publicly that roughly ten percent of its visuals were AI-generated, made with Meta's tools, and took the argument that followed on the chin. People disagreed about the choice. Nobody could accuse him of hiding it. Disclosure doesn't win you the argument, it wins you the right to have it, and that's the whole point. Deepfakes and synthetic likeness sit on the same fault line between spectacle and trust that's reshaping narrative film, and I've written separately about where that's taking Hollywood.

Lines I won't cross

Some things don't get better with a label. These are mine, and they're consistent with my policy whatever the tool, the deadline or the budget.

  • A testimonial that didn't happen, however assembled

  • A founder or leadership message in a cloned voice without written, use-limited consent

  • Product-in-use or location footage presented as captured when it was generated

  • Anyone's face or voice without consent that specifically covers synthetic use

  • Charity, documentary or evidential material presented as observed when it wasn't

  • A label used as a cure for a claim the pictures can't support

That last one matters most, because it's where well-meaning teams go wrong. Disclosure protects an honest piece of work. It doesn't launder a dishonest one, and the ASA says the same.

What you should actually be filing

Every job that touches the altered or synthetic tiers generates five pieces of paper, and none of them takes long.

  • Consent that names synthetic use, not just "appears on camera"

  • A one-page AI log listing the tool, what it touched, who signed it off and when

  • The exact disclosure wording used, and where it appeared

  • Source files for anything presented as real

  • The tier you'd declare if a fund, festival, insurer or buyer asked tomorrow

If you commission rather than produce, the same list becomes your delivery requirement, and there's a fuller version in my guide to commissioning AI-assisted video safely.

Where the risk concentrates

A few situations deserve more caution than the general rules suggest. Your own employees and founders, because the power imbalance means a hallway yes isn't consent, which is the whole Sally problem. Customers and testimonials, because belief in them is the product. Children, where consent needs to be separate, tighter than adult talent terms, and a parent being present doesn't authorise a generated face or voice. The dead, where an estate's permission and the audience's trust are two different things, which is exactly what Roadrunner ran into. And archive presented as found, where generation and reconstruction need the same care and the same label, which is what What Jennifer Did skipped.

Bias is a production risk, not a talking point

AI systems inherit the gaps in what they were trained on, and those gaps surface in outputs that look polished. The generic advice is to review for bias. The production version is more specific. Who was represented in the reference material, and which accents, ages, skin tones or disabilities is the tool thin on. Did it introduce stereotypes while "completing" a scene. Was the output reviewed by someone with relevant lived or professional experience rather than by the team that made it. And if a group is affected by how they're shown, did anyone give them a meaningful chance to object before it shipped. Where AI touches personal data, the ICO's guidance on fairness and accountability in AI is the UK reference to align with.

Pressure points that catch real teams out

Ethics often collapses under pressure, not because teams are careless, but because real production is full of awkward moments. These come up a lot.

A marketing team wants a synthetic founder voiceover. If a founder's voice is being cloned, you're dealing with identity and trust. Treat it as high risk. Get explicit consent in writing, agree where and how it can be used, and build a sign-off trail. If you can't document that, don't ship it. Cloned voices raise the same questions for voice actors and the businesses that hire them, and the answer keeps coming back to consent and scope.

A client supplies a viral clip and asks you to repurpose it. This is where provenance matters. Before you edit or publish, verify the source, check rights, and sanity-check whether it could be manipulated. If you can't confirm where it came from, it's safer to walk away than to become the distribution point for misinformation. Industry work on content provenance is what standards like C2PA are for, tracing the origin and edits of a piece of media.

A vendor won't explain their AI pipeline. This is a quiet red flag. You don't need trade secrets, but you do need clarity on consent, rights and ownership. If a supplier can't describe how they handle likeness and voice consent, or who owns the outputs, treat it as a risk and choose another route.

Five questions to ask any AI video vendor

If you outsource any part of AI-assisted production, ask five questions. The aim isn't to interrogate. It's to protect your brand.

  • What tools are being used in the pipeline

  • Who owns the outputs, and can they be reused elsewhere

  • How is consent handled for voice, face, and likeness

  • What is the approval process and can you evidence it

  • What happens if a claim is challenged later

A good supplier will answer these without defensiveness. If the answers are vague, that's useful information.

A working model for AI in production

Concept Benefit Application
Consent as infrastructure Protects identity, reputation, and legal position Document explicit approval for any synthetic face, voice, or likeness
Disclosure as trust protection Prevents audience confusion and reputational harm Disclose when AI alters who appears to speak or what appears to happen
Bias as production risk Avoids tone-deaf or exclusionary content Audit outputs with reviewers from outside the core production team
Provenance checks Reduces the risk of spreading manipulated or stolen media Verify sources and rights before repurposing third-party clips
AI as workflow support Speeds production without undermining authenticity Use AI for cleanup, captions, and fixes rather than synthetic presenters
Detection as triage Flags suspicious content without over-reliance Use detection tools as a first pass, not a final verdict

Detection tools won't settle it

There are scanners that flag suspected synthetic media, and they have their place as a first pass. What they can't do is prove something is authentic or ethical, because false positives and false negatives are still routine, especially on compressed or heavily edited video. Provenance records, consent paperwork and editorial judgement do the real work. I've gone deeper on detection, and on why trust breaks when context gets stripped away, in a separate piece on AI video and trust.

What this page is and isn't

This isn't a legal manual and it isn't a technical guide to spotting synthetic media. It's an editorial and production standard for one question, which is when AI use changes what an audience may reasonably believe, and what to do about it when it does. The rules move, so check the linked sources on the day, and where I don't know something, like where the UK will land on digital replicas, I've said so rather than guessed.

Questions people ask

How do I use AI ethically for video generation?

Run every use through four tests. Has the tool changed who appears to speak, what they appear to say, what appears to have happened, or whether a person shown is real, present and consenting. If yes to any, you need consent from anyone represented and a visible label. If no, it's workflow, and workflow doesn't need announcing.

Which AI video tools are the most ethical to use?

No tool is ethical on its own, the use is. Tools do differ in how they were built, so ask four things. What the model was trained on and whether the vendor will confirm it in writing, whether output carries Content Credentials, what happens to material you upload, and whether it supports the labelling you're obliged to do.

Do I have to disclose every time AI touches the timeline?

No. Transcription, captions, noise reduction, colour work and routine clean-up don't change what anyone said or did, and labelling them empties the word for when it matters. Disclose when AI changes the substance of what a viewer would believe, which is what the four tests are for.

Does UK advertising law require an AI label on ads?

No. CAP guidance from May 2025 says there's no blanket duty to disclose AI in UK ads, and the codes apply however an ad was made. The real question is whether the audience would be misled without a label, and a disclosure won't rescue an ad that's misleading anyway.

When does EU AI Act Article 50 apply to a video I'm making?

From 2 August 2026, and the cut-off is the date of generation, so material generated earlier doesn't need retro-labelling. Work seen in the EU can be in scope even if you didn't target the EU, and the labelling duty follows whoever controls the AI use, which is worth settling in the contract.

Is a cloned voiceover okay if the person agreed in a meeting?

A verbal yes isn't enough, and with your own staff the power imbalance makes it worth even less. You need written consent naming the permitted uses, a defined end date, and an agreement about what happens to the voice model if the person leaves. Without that, don't ship it.

Where should the label go, on screen, in the description, or in the file?

It depends what could mislead. On-screen labels at the moment of the synthetic shot are strongest. Descriptions and credits are the documentary convention. Platform toggles are required regardless on YouTube, TikTok and Meta. Content Credentials in the file help but often don't survive an export, and delivery paperwork covers funds, festivals and buyers.

What are the biggest ethical and safety risks with AI video, and how do I avoid misuse?

The recurring failures are missing consent for a real person's face or voice, synthetic material presented as real, testimony that was altered or manufactured, and provenance nobody can evidence. The protection is boringly practical. Written consent, a tool log, a visible label where the four tests say so, and source files kept.

What do the industry guidelines for AI in film and TV actually say?

The APA's documentary guidelines say generative material must never pose as primary source, and that AI belongs on cue sheets and in credits. The BFI's screen sector report recommends public transparency, citing strong audience support for disclosure. Human Provenance in Film proposes declared tiers, and performer bodies focus on consent, compensation and control.

What should marketing teams look for in a video production partner that uses AI responsibly?

A partner who'll tell you which tools touch your job, who owns the outputs, how consent for faces and voices is documented, what the approval trail looks like, and what happens if a claim is challenged. Straight answers to those five questions matter more than any badge or portfolio piece.

Read Next

If you're commissioning or publishingThe practical next steps from here

Go deeperWhere this is all heading

What changed and when

September 2026. Full rebuild. New title and opening, the three-tier map with Human Provenance in Film equivalents, the rules table checked against primary sources, the industry guidelines section, three documented cases, red lines, the filing list, and this FAQ. The four tests, the minimum standard, the pressure points, the vendor questions and the working model are unchanged from earlier versions of this page.

Primary sources are linked where each claim is made. Rules were last checked on 17 September 2026.

Nigel Camp

Filmmaker and author of The Video Effect

Previous

ASMR video production for brands
Next

Will AI Replace Human Video Editors by 2030?