Real or AI-generated video? How audiences decide what to trust
Updated 17 September 2026
AI-generated video is not automatically untrustworthy. The trust problem starts when synthetic or altered material is presented as captured reality, especially when it speaks for a real person, demonstrates a real product, or claims to document a real event. So the question I ask of every piece is not "was this made with AI?" It is "what is the viewer being asked to believe?" If the answer is that something really happened, capture it or say exactly what changed.
I have made films for brands and charities for fifteen years, mostly with real people, real lenses and real light. I have only recently started working with generated video myself, and that is part of why I am careful about where the line sits. This page sets out how I answer the question I now hear most often from people who commission video. When is real footage worth it, when is generated footage fine, and how do you keep an audience's trust either way?
Picture a situation the newest editing tools now make easy. A charity films an interview with a volunteer in a community kitchen. In the edit, one cutaway runs two seconds short, so the editor uses a tool that extends the shot. The generated seconds show the volunteer handing a plate across the counter and smiling. It looks seamless. Then the volunteer watches the finished film and says, quietly, "I don't remember doing that."
Nobody set out to deceive anyone. The tool worked exactly as designed. But a real person is now on screen doing something they never did, in a film that asks its audience to believe it is real, and the first person to lose trust in it is the one who was filmed. The tool did not fail. The process did, because nobody stopped to ask what the viewer, or the person on screen, was being asked to believe. That question is what this page is about.
Explore this guide
If there is one section to start with, begin with How viewers decide whether a video is real. It explains the small moments that make people pull away before they realise it.
How viewers decide whether a video is real
Most viewers do not check a video for AI. They react to it, often in under a second and often before they could say why. That is my reading of the research below, and it matches what I see in my own work. It changes the useful question from "is this AI?" to "does this still feel answerable to a person?"
A short clip slid past in someone's feed. A young woman dancing in her kitchen, half laughing at herself. Steam lifted from a kettle behind her. It felt like a real moment, and that was the point, because generated video now lives in the same feeds where people expect unguarded human moments.
Then a commenter pointed out the tell. For a fraction of a second, her hand did not meet the cupboard door. It passed through it.
The mood shifted, not into outrage, but into distance. The clip still looked lively and competently made, but it no longer felt like a person letting you in. It felt like output. And that doubt rarely stays with one post. Viewers carry it into the next ad, the next founder clip, the next charity appeal.
Trust does not break with a bang. It breaks with a pause. The moment you think, what am I watching, the story is no longer holding you. You are holding the story at arm's length. And in a flood of clips, most people do not argue. They leave. You do not get a comment saying I no longer trust you. You just stop being watched.
The most quoted number of 2026 is that 83% of consumers can spot AI video. That is not what the study found. Animoto's State of Video report, a vendor survey of US consumers and marketers, found that 83% had watched a video they suspected was AI-generated. Suspicion, not detection. The same survey found 36% say an AI-generated brand video would lower their view of the brand, while over a third trust AI content as much as content made without it. The audience is split, not uniformly hostile. A much larger, independent survey of US adults by the Pew Research Center found the same tension from the other side. 76% say it is extremely or very important to be able to tell whether media was made by AI, and 53% are not confident they can.
My reading of those numbers is simple. People want to know, and they know they cannot reliably tell. Suspicion without skill is the default posture now, and a viewer who suspects is a viewer who has stopped relaxing into the story. I felt this directly when a client's first reaction to a human voiceover I had booked was to ask whether it was AI. I have written about that, and about where AI has and has not taken voice work, separately.
This is not only about AI, either. Editing alone can change a narrative. Two shots swapped. A reaction pulled from a different moment. A pause cut out. If the cut nudges people towards a conclusion without them noticing, trust does not simply drop on that video. It drops on the creator, the brand, and sometimes the subject too.
The check I use has three parts, in this order. Disclosure asks whether you have told people, plainly, at the moment it matters. Provenance asks whether the origin still makes sense after reposts, crops and compression. Accountability asks whether a named person will stand behind the work and answer for it.
The check I use has three parts, and the order matters.
Disclosure
Have you told people, plainly, at the moment it matters?
Goes first, when the caption is strippedProvenance
Does where it came from still make sense after reposts, crops and compression?
Blurs next, as the file travelsAccountability
Will a named person stand behind the work and answer for it?
The part that has to hold when the others failIf nobody is accountable once the caption and the context have gone, the viewer does the only thing left and withdraws. What happens after the context falls away is its own problem, covered in why AI video trust breaks when context is stripped away. The stakes rise fastest where synthetic media starts behaving like evidence. That happens in live broadcasts and events, where people read generated visuals in real time before anyone can add context, and wherever a recognisable face or voice can be faked convincingly.
Assisted, altered or synthetic. What the terms mean
Most arguments about AI video go wrong because people use "AI video" to mean three different things. My AI policy uses three tiers, and I use them here too.
| Tier | What it means | Examples |
|---|---|---|
| Assisted | AI helped with the work but did not change what the video claims. | Captions, noise reduction, reframing, rough assembly, colour matching |
| Altered | Real footage or audio was materially changed. | Face replacement, de-ageing, a re-voiced line, a generated background behind a real person |
| Synthetic | The image, voice, performance or scene was created by a model. | A generated shot, a synthetic presenter, a cloned voice, a fully generated film |
Two terms come up often around these. A digital replica or digital double is a generated or altered version of a real person's face, voice or performance. Provenance is the record of where a piece of media came from and what was done to it.
The middle tier is where most trouble happens. Assisted work rarely misleads anyone. Openly synthetic work usually does not either. Altered work is where real footage borrows its credibility for something that did not happen the way it looks. The film industry is drawing similar lines. Human Provenance in Film, a disclosure standard launched at Cannes in May 2026 by the London sales agency The Mise En Scène Company, uses three tiers of its own, No AI Used, Assistive AI and Generative AI, and its consultation is open until 31 October 2026. Its tiers line up with mine rather than matching them exactly.
Why perfect footage can start to feel fake
Perfect footage stops reassuring people once perfection is cheap. The first wave of synthetic video won attention by being impressive. The next wave risks losing it by being frictionless, not because viewers become purists, but because polish stops working as a sign of care once anyone can buy it.
When every face is perfectly lit, every delivery perfectly paced and every reaction feels calibrated, many viewers start to sense the smoothing itself. They may keep watching, but in my experience they invest less. A performance produced without risk, fatigue, or a lived past behind the eyes can read as technically correct and emotionally weightless. The giveaways people name in the Animoto survey fit that. Robotic gestures, unnatural voices, a lack of emotional tone.
Viewers rarely rage about it. More often they disengage, and disengagement leaves no trace. Generated content is here to stay, and the risk is not that it exists. The risk is volume, because volume makes everything feel weightless. So the goal is not to look perfect. It is to stay believable.
That is why post-production matters more now, not less. When polish is cheap, value moves into the finishing choices that keep texture, restraint and judgement. In September 2026 Adobe put a tool inside the Premiere timeline that generates a shot to fill a gap in your cut. The temptation I used to describe is now a button. The same flattening can start earlier than the edit, when drafting gets faster and intent gets thinner, which is the question behind what stays human in the script. In post, I have written about what the new tools do to the editing job by 2030, and why the next era will be won pixel by pixel in post.
When real footage is worth more
Real footage is worth more when the video is asking to be believed. When perfect becomes cheap, credible reality becomes valuable, and you can already see an authenticity premium forming, where attention sticks to work that signals a real person took a risk, made choices and stood behind them. That is not about mess for its own sake, and a real image is not automatically a trustworthy one. It is about the viewer sensing lived effort rather than manufactured ease.
Audiences are saying something similar. Research cited by Human Provenance in Film found 77% of consumers want to know whether content was made by AI in whole or in part, and 70% would rather watch a film or show made by a human, drawing on Baringa and Deloitte. Film workers interviewed for a peer-reviewed study during the Hollywood strikes kept reaching for the same word for generated output, soulless, and tied what they do to lived experience rather than technique. That is a small qualitative study of workers, not audiences, but it names the thing audiences seem to be paying for.
This will not play out the same way everywhere. Some work should choose speed and volume, and that is a legitimate choice. The honest version of the premium is narrower. Real footage is the wrong spend when the video is not asking to be believed.
The same shift is changing how performers are valued. Once identity can be copied, permission becomes part of the meaning, which is why actors are starting to look like royalty-earning IP. It is also why consent runs through this whole page, and why using someone's likeness or voice in AI video gets its own piece.
When should a brand use AI-generated video?
Use AI-generated video when it is expressing an idea rather than proving that something happened. Synthetic visuals can work well for concepts, explainers and openly fictional creative. Use real footage, or disclose the alteration clearly, when viewers may treat the video as evidence of a person, a product, a place or an event.
The test I use is one question. What is the viewer being asked to believe? The closer a video gets to functioning as evidence, the more provenance, disclosure and consent matter. That principle survives whatever the tools do next.
| The video | What the viewer assumes | My default |
|---|---|---|
| Customer testimonial | This real person said this | Use the real speaker and their real words. A synthetic reconstruction can still mislead even with a label, because it reads as an actual endorsement. |
| Founder or leadership message | This person is speaking to me | Real. The video exists to show this person said this. A synthetic version undermines its only job. |
| Product demonstration | The product behaves this way | Real, or clearly illustrative. Generation is not the problem. Implying the product did something it has not been shown to do is. |
| Charity appeal or documentary | This happened, to these people | Real, with disclosed exceptions. Composites and re-voicing to protect a source can be right, and must be said out loud. |
| Explainer or concept visuals | This illustrates an idea | Either. Nobody reads an animated diagram as evidence. Generation is often better value. |
| Openly synthetic creative | This is meant to be artificial | Generated is fine when the artificial nature is clear from the context, not only from a small label. |
| Volume social content | This is promotional creative | Either, with care. Watch for synthetic faces or scenes that create an impression the brand cannot stand behind. |
In the UK the advertising rules already cover this. The regulator's position is that there is no blanket duty to disclose AI in ads, the same rules apply however an ad was made, and disclosure is very unlikely to fix a message that is misleading in the first place. It has already acted on this. In 2025 it ruled against a boot seller whose ads claimed the boots were handmade by "Henry" in his workshop, alongside images of a craftsman the regulator understood to be AI-generated. There was no evidence the boots were handmade by Henry, and the claims were found misleading. The lesson is the one in the table. The problem was the impression, not the software.
What I would not generate
A customer testimonial
A founder or leadership message
Product-in-use footage presented as captured
Anyone's face or voice without their written consent
Charity or documentary evidence presented as observed
If you are the one signing off the brief, how to commission AI-assisted video safely covers the questions to settle before anything is made.
Access is widening fast. Tools that once sat behind agency budgets are now open to local brands and solo operators, but access does not solve trust, and where AI actually fits for small teams is a question of judgement more than budget. How the viewing context changes what a piece can get away with is covered in where video lives now.
How to prove what happened before you publish
You prove a video before you publish, not after someone asks. Proof is a bundle of small signals, and they fall into three areas. Provenance, meaning how the work was made. Proof it happened, meaning what existed beyond the edit. And intent, meaning why the choices were made and who answers for them. Tools can help with provenance, but they cannot supply intent. Labels help with disclosure, but they do not prove anything once a clip is reposted without them.
The technology is moving. Content Credentials, the C2PA standard for attaching a file's history to the file, moved from idea towards infrastructure in 2026. In May, OpenAI joined the C2PA steering committee and began adding Google's SynthID watermark to its generated images, and in July extended it to audio. Coverage of video is still uneven, so do not assume a generated clip carries a reliable mark.
Platforms are moving faster than the law. Since May 2026, YouTube automatically labels videos where its systems detect significant photorealistic AI, even if the uploader said nothing, and shows the label under the player or on the Short itself. That is useful on YouTube. It means nothing once the file is downloaded and posted somewhere else. Design for the version with the caption stripped.
The rules are catching up too. For EU-facing work, Article 50 of the AI Act has applied since 2 August 2026. Broadly, toolmakers must build in machine-readable marking of synthetic content, and whoever publishes a deepfake, meaning generated or manipulated content that could pass as a real person, place or event, must disclose it. Evidently artistic or fictional work has lighter treatment, and the exact duty depends on your role and the content. The UK position, which rests on the advertising rules above and on data protection where real people are involved, is set out with the sources in my AI policy.
None of that replaces judgement, consent or accountability. If real people, likenesses, voices or personal data are involved, the ICO's guidance on AI and data protection is the baseline. This is a working editorial position, not legal advice, and rules vary by territory and by your role in publishing the work. My own routine for labels, consent and records is in my AI policy. What follows is the viewer's side of it, the checks that decide whether a piece will hold.
Before anything goes live, I run a 30 second check and answer it like a viewer, not a producer. What changed? What exists beyond the frame? Who stands behind the choice? If you cannot answer cleanly, the viewer will answer for you.
Five habits worth stealing
Name what changed and what did not. Specific beats vague.
Treat faces and voices as consent-first material.
Keep originals when claims matter. Rushes, audio, project files.
Assume labels fail once content is reposted.
Protect the spell. Do not turn your viewer into an investigator.
Disclosure does not have to ruin the mood. If AI has changed something that affects what a viewer thinks they are seeing, hearing or inferring, say so before anyone starts playing detective. A small upfront note lets the viewer exhale and stay with you. You do not need to announce every technical step. You do need to be clear about synthetic voice, generated performance, composite scenes, or clean-up that changes what looks like direct evidence. Disclosure stops the viewer being misled about what they are seeing. It does not make an unauthorised likeness, a fabricated testimonial or a false product claim acceptable. The pre-publish test is one line. Can a reasonable viewer tell what kind of truth claim this video is making?
Copy-ready disclosure lines
Keep them short and put them where people actually look.
| Scenario | Disclosure line | Best placed |
|---|---|---|
| Light AI clean up (Assisted) | AI used for clean up and pacing. No performance was generated. | Description, end card, or credits |
| Blended elements (Altered) | This film contains AI-generated imagery where indicated. Everything presented as real footage is real footage. | Caption, description, or pinned comment |
| Voice cloning (Altered or Synthetic) | Voice cloned with permission. Disclosure is part of consent. | Description, pinned comment, or credits |
| Digital double or de-ageing (Altered) | A performer's likeness was digitally altered with their consent. | Credits and description |
| Fully generated (Synthetic) | This film is AI-generated. No footage was captured. | On screen at the start, and in the description |
Here is how that works in practice. A documentary uses a composite character to protect a real source, and the narration is voiced with AI, with permission. Nothing is trying to pass as a raw recording, but viewers still deserve to know what they are hearing. The voice cloning line goes in the description or pinned comment, so nobody starts wondering what else was changed. For the paperwork behind all this, permissions, records and the proof trail you may need later, who owns the IP and how you prove you made it goes further. And because finished work can now be quietly revised, what can be changed after a film is released belongs in the same conversation.
Questions people ask about real and AI-generated video
Can you trust AI-generated video?
It depends on what the video is asking you to believe. AI-generated footage used openly as creative illustration is a different thing from synthetic footage presented as evidence of a real person, product or event. Trust depends on the context, on clear disclosure, and on someone being able to explain how the material was made.
Can people tell if a video is AI-generated?
Not reliably. In a large US survey, most adults said it matters to know whether media was made by AI, and more than half were not confident they could tell. Viewers often sense something is off, such as robotic movement or a flat voice, without being able to name it. That is why disclosure matters more than detection.
Should my brand use AI-generated video?
Sometimes. It can be the right choice for explainers, concepts, fictional creative and openly synthetic work. Take more care when the video could be read as evidence of a person, a product, a place or an event. Testimonials and founder messages should normally use the real person and their real words.
Does a label make an AI-generated video trustworthy?
Not on its own. A label tells the viewer what kind of thing they are watching, and that keeps them from feeling misled. It does not make a fabricated testimonial, an unauthorised likeness or a false product claim acceptable, and it does not travel with the clip once it is reposted without the caption. Whether you are required to label, in the UK and for EU audiences, and the wording and placement I use, is set out in my AI policy.
Do I have to disclose AI video on YouTube?
Yes, for realistic content that could pass as a real person, place or event, YouTube requires creators to disclose it at upload. Since May 2026 YouTube also applies its own label when its systems detect significant photorealistic AI, even if the creator did not disclose it. That label stays on YouTube. It does not travel with the file if the clip is downloaded and reposted.
Is an AI product demo legal in the UK?
Generation itself is not the problem. Presenting a generated demo as footage of the product working is. UK advertising rules apply however an ad was made, so if a reasonable viewer would take the clip as evidence of what the product does, it needs to be accurate and backed up. A disclosure is very unlikely to cure a misleading impression.
What is the difference between AI-assisted, AI-altered and AI-generated video?
AI-assisted means AI helped with the work without changing what the video claims, such as captions or clean-up. AI-altered means real footage or audio was materially changed, such as a replaced face or a re-voiced line. AI-generated means the image, voice or scene was created by a model. The altered middle tier is where most disclosure failures happen.
Is real footage worth the extra cost?
When the video is asking to be believed, usually yes. A real person, place and moment give the viewer something to trust, and that trust transfers to the brand. When the video is illustrating an idea rather than proving something happened, the case for real footage is weaker and the budget may be better spent elsewhere.
Where to go next
The signal that survives the flood
The future of AI filmmaking will not be decided by who generates the cleanest surface. It is more likely to be decided by who keeps human judgement visible when synthetic capability is cheap and everywhere.
That is the real battle for the soul of this work. Not human versus machine. Clear authorship versus blurry accountability. When output is abundant, what viewers keep paying attention to is the sense that someone meant this, chose this, and will stand behind it.
Images on this page are AI-generated illustrations.
