The case for keeping real footage: what an agency should never fake with AI

Most advice about AI video is about what you can make. This page is about what you should not make, because an agency carries the risk when a fake slips into a client's campaign. If you are deciding where AI video belongs in your studio at all, start with the comparison of AI video, stock footage and filming it yourself. This page is the stop list that sits on top of that decision.

Nothing here is legal advice. Rules differ by country, by industry and by platform, and they change. Where I name a rule I point to the official source, and you should check local law and the platform's current rules before you rely on any of it.

The short version: the never-fake list

I would put these six lines in the agency handbook and in every statement of work (the document that defines what you are selling a client).

1. Testimonials and reviews. A customer who does not exist, or who never said the words, is not a testimonial.
2. Results and before and afters of real work. If the picture claims "this is what we did", it must be what you did.
3. The physical state of a property, product, venue or person. The room, the dent, the queue, the staff.
4. Named or recognisable real people saying or doing anything without their written agreement.
5. Regulated claims shown as outcomes. Health, money and legal results.
6. News-like footage of events that did not happen.

In each case the viewer is asked to treat the video as evidence, and evidence needs a real camera, a real person or a real record.

Testimonials and reviews

This is the line with the clearest official guidance. In August 2024 the FTC announced a final rule banning fake reviews and testimonials. Its announcement names AI-generated fake reviews and says the rule lets the FTC seek civil penalties against people who knowingly break it. I am not stating a penalty amount or an effective date, because I could not confirm either from the FTC's own announcement page. Look both up on ftc.gov before you quote them to a client.

The FTC's separate question-and-answer page on endorsements is useful for the everyday cases. It says an endorsement must reflect the honest opinion of the endorser, and an endorser cannot describe an experience they have not had. An AI person who "loved the service" has never had the service. That is the whole problem in one sentence.

The rule also matters to you as the producer. A rule that reaches people who knew or should have known something is fake does not leave much room for "the client asked for it" as a defence. I would raise who carries the risk with a lawyer, in the contract, before the work starts. I cover contract wording in the AI clause for agency contracts.

What you can do with a real client's real words

There is a narrower, honest use. A real customer gives a real testimonial, in writing or on camera. You may want to re-voice it, translate it, or turn it into an illustrated video. A legal-marketing vendor's guide on this (a vendor, so treat it as one view, not law) suggests the same pattern I would use:

• Written consent from the customer that explicitly mentions AI generation.
• The customer sees and approves the final cut.
• No added claims. Their words, nothing extra.
• Consent is confirmed again before the clip is used in paid ads.

If any of those four is missing, shoot it properly or do not use it. An invented person is never a testimonial, whatever you call it.

The avatar edge case

An AI presenter reading a script is not a testimonial. Brands use them for training, explainers and product walkthroughs, and I cover that in the AI spokesperson guide. The risk starts when the avatar is presented as a real employee, a real customer or a real expert. Either say it is a virtual presenter or use a clearly illustrated character. If a real face is part of why people trust the brand, read the page on giving a faceless brand a human face before you replace it.

Results, case studies and before and afters of real work

A before and after makes a promise: this is what happened. The FTC's endorsement guidance adds a second layer. Where a result is exceptional, you need to show what is typical or say plainly what a customer can generally expect. That rule applies whether the picture was shot or generated, and generated pictures make it easier to cheat without noticing.

Here are the cases I would keep human:

• Your own agency case studies. Screenshots of real dashboards, real client logos used with permission, real numbers. An AI-made "result" chart is a fabricated record.
• Client before and after photos. A renovation, a landscaping job, a skin or fitness result, a repainted car. If the viewer will use the picture to judge what they will get, it has to be the actual work.
• Anything with a number on screen. A claim of "doubled leads" needs a source you can show, not a render.

There is a safe version. A clearly labelled illustration of a process ("how a renovation moves from plan to finish") is explanation, not evidence. Label it as an illustration and keep it away from the results section of the page.

A made-up example

Say a landscaping client asks for a "garden transformation" video. You have a real before photo and the finished garden is only half planted. The temptation is to generate the finished version. I would shoot the half-planted garden honestly, show the real before, and sell the plan as an illustration marked "design concept". Then film the finished garden next season. That is slower, and it is the version that survives a complaint.

Property, product and venue condition

The viewer assumes the thing in the video is the thing they will receive or visit. AI video can change that without anyone intending to: a larger room, a nicer view, a cleaner finish, a quieter street, a product that looks better than it is.

Four checks I would make on any shot of a real place or product:

1. Is the geometry right? Room size, window count, ceiling height, distance to the water.
2. Are the materials right? Flooring, fittings, finish, colour under real light.
3. Is anything added or removed that changes the decision? A balcony, a power line, a damaged wall.
4. Would the owner sign off that the video shows what the buyer will find?

Property is the strictest case, and some local real-estate rules require a disclosure for edited listing images. Where the video sells a specific thing, I would film that thing. Where it sells a mood around the thing (a lifestyle scene next to the real footage), generated shots can work if they are marked as illustrative and sit apart from the property tour. See the real estate video tools comparison.

Real people: faces, voices and words

Putting a real person in a generated video, or cloning a voice, raises consent and publicity-rights questions, and several platforms also restrict it. A person's agreement should be written, specific to AI use, and specific to the campaign and the length of time. The terms of the tools themselves point the same way. Higgsfield (a web platform that hosts many image and video AI models) says in its Terms of Use that you warrant you hold the rights and consents for any name, likeness or voice you upload, and it bars impersonation. That is a tool vendor pushing the risk onto the user, which is a good reason to settle consent before you open the tool.

For staff, customers, founders and talent, my rule is simple: no face and no voice without a signed release that mentions AI. For anyone else, no. That includes celebrities, politicians and competitors.

Medical, money and legal claims

Health, money and legal advertising is where regulators ask for proof before you claim anything. A generated patient who "got her energy back", a generated investor who "doubled his savings", or a generated client who "won his case" shows an outcome that no real person had. Even with a small-print label, it is a claim you cannot substantiate.

What I would do instead for clients in these fields:

• Use real, consented people, with the typical-results wording the relevant regulator asks for.
• Use animation or illustration for explanation: how a procedure works, how a pension is taxed, how a claim is filed. Mark it as an explainer.
• Send the script and the final cut to the client's compliance contact, and keep their approval on file.

Platform rules: label what a viewer could mistake for real

Platform labelling is a switch, not a moral stance. You set the toggle that matches the content, and you keep a record.

YouTube's help page is the clearest official source on this. It asks creators to disclose realistic altered or synthetic content, for example a real person appearing to say or do something they did not, an altered real event or place, or a realistic scene that did not occur. It says minor edits and clearly unrealistic content do not need the disclosure, and it warns that repeated non-disclosure can lead to labels, removal or suspension from its partner programme.

I am not describing Meta's or TikTok's rules here, because both have AI labelling policies that change often and I cannot give you a verified summary. Open each platform's current ad and content policy before a campaign goes live. The same is true of marketplace and ad-network rules, which I also leave to the official pages.

One more point. Higgsfield's terms say it may embed machine-readable provenance signals, such as ones based on the C2PA (Content Credentials) standard, does not promise they will be applied or will persist, and tell users not to remove them and to disclose AI generation where the law requires it. In other words, do not count on the tool to label your content for you. See whether agencies should tell clients and audiences about AI use for the disclosure habits side.

Law at a glance (information only)

Source	What it says in plain words	Hedge
FTC final rule on fake reviews and testimonials (2024)	Bans fake reviews and testimonials, names AI-generated ones, civil penalties possible for knowing violations	Dates and amounts not confirmed here, check ftc.gov
FTC Endorsement Guides Q and A	Honest opinion, real experience, typical results or a clear statement of what to expect	US only
EU AI Act, Article 50	People who deploy AI that makes realistic image, audio or video of real people, places or events must disclose it; lighter duty for evidently artistic or fictional work	Applies from 2 August 2026; the Digital Omnibus did not delay it for people who make the content. Information, not legal advice
YouTube altered or synthetic content policy	Label realistic synthetic content, skip it for clearly unrealistic content or minor edits	Policy can change

Local advertising codes may add more. This table is a starting point for a lawyer conversation, not an answer.

A safer-use list and a three-question test

I say "safer", not "safe". Nothing is safe everywhere.

• Abstract backgrounds, textures and motion graphics.
• Concept boards, storyboards and pitch mockups, labelled as concept (see mood video for a client pitch).
• Animated explainers and invented characters that nobody mistakes for a real customer.
• Imaginary scenes and fantasy settings for a product, in an ad that reads as creative.
• B-roll of generic things (clouds, traffic, a plant growing) that the ad does not claim as the client's own.
• Internal drafts that are replaced before they ship.

Licensed stock footage can be the safer pick for neutral b-roll when a client's legal team wants clear provenance. It comes with a licence and a source.

The three-question test, per shot

1. Will the viewer take this as a record of something real?
2. Could a reasonable viewer be misled about the product, the result or the person?
3. Is a real person's face, voice or words involved?

Any "yes" means one of three things: shoot it, license it, or label it and get written sign-off from the client. Add one habit: keep a log of which assets were AI-made and which platform toggle you set. It is a ten-second note per asset and it answers most later questions. For the rights side of commercial use, see can I use AI video commercially.

What to put in your contract and your client note

Here is a clause you can adapt. Have your own lawyer review it, because it is a starting point and not legal advice.

AI-generated material. The Agency will not use AI-generated or
AI-altered images, video, audio or text to depict (a) testimonials,
reviews or statements attributed to any real or invented customer,
(b) results, case studies or before-and-after outcomes of work
actually performed, (c) the physical condition of any property,
product, venue or person, or (d) any identifiable real person, unless
the Client and, where applicable, that person have given prior written
approval that specifically refers to AI generation. Any AI-generated
material in a deliverable will be identified in the AI use note.
The Client is responsible for the accuracy of claims it supplies and
for providing substantiation on request.

Then ship an AI use note with every deliverable. Keep it to a few lines:

AI use note. Asset: [name]. Tool: [tool and model]. Synthetic
elements: [list]. Real footage: [list]. Platform label set: [yes/no,
which]. Client approval: [name, date].

The quality side of this, the checks you run before you send, is in the AI video quality check.

When the client insists

It happens. A client wants a testimonial video and has no testimonials, or wants "a real-looking customer". My answer is a short script: "I can make you an illustrated explainer or a clearly labelled concept. I cannot make a customer who does not exist say they used you. Here is how we get a real one, fast." Then offer a cheap way to a real one: a short phone interview with a happy customer, a signed release, and a light edit. Often the real version takes a day and costs less than the argument.

If the client still insists in writing, I would stop and take advice. An email asking an agency to fabricate a testimonial is not something to ignore.

Who should skip AI video altogether

• Agencies whose clients sell in medicine, finance or law and who have no compliance reviewer on the client side.
• Any job where the main asset is a person's trust, such as a founder story, a local trades business or a clinic.
• Work for a client who will not sign an approval, or who will not let you label the content.

Sort work into "evidence" and "illustration", and keep the camera for evidence.

How I'd approach this

I would write the six-line never-fake list, add the clause and the AI use note to the standard contract, and run the three-question test on every shot at the storyboard stage, when changing course costs nothing. Put the approvals in one place so a client can see who signed off what. If you want the repeatable version of that, with an approval portal and a shared asset library around it, that is the kind of workflow Blur Studio builds for agencies. Whatever you use, keep the camera for anything the viewer will treat as proof.