Disclosing AI-Generated Media to Your Audience

Disclosure conversations usually start in the wrong place: with a rule, and an argument about whether this particular video falls inside it. The rules are worth knowing and they are not the interesting part, because they change and because they set a floor rather than a standard.

The question that actually decides whether an audience keeps trusting you is narrower and more stable. Are they being asked to believe something about a person, a place, or an event that generation has made untrue? If yes, tell them. If no, a label is noise. This guide covers how to apply that test, where the genuinely difficult cases sit, and how to word a disclosure so it informs rather than hedges.

The Test

Ask what the viewer is being invited to believe.

A generated shot of clouds moving over a city, used as texture behind narration, invites no belief about anything. Nobody watching it concludes a specific thing happened at a specific place. It is illustration, in the same category as a stock landscape or a drawing.

A generated shot of a person saying words invites a very specific belief: that this person said this. A generated depiction of an event — a protest, a product failure, a weather incident — invites the belief that it occurred. Both of those are claims, and generation makes them false.

Between those poles the test still works, applied honestly. The question is not "did I use a tool" but "is the audience forming an impression that is not true, and would they mind".

The row that people argue about is the third. A generated street scene used to illustrate "cities are getting busier" is illustration. The same shot cut into a news-style piece about a particular city becomes evidence, and the change happened in the edit rather than in the generation.

Process Versus Product

A distinction that clears up most of the anxiety: disclosure is about what the finished content depicts, not about which tools touched it.

Nobody expects a note explaining that the audio was denoised, the transcript was automatic, the footage was upscaled, or the colour was matched by software. These are production processes, and disclosing them tells the audience nothing they can use.

Generated depictions are different because they change what the content asserts. The line falls between tools that helped you present something real and tools that produced something that was not there.

This also explains why the answer differs between two videos that used identical software. The tool is the same; what it produced, and what the audience takes from it, is not.

The Hard Cases

Three situations resist the test and deserve thinking about in advance.

A synthetic narrator on factual content. The narration is generated; everything it says is true and verified. Nobody is being deceived about the facts, and the audience may still feel differently about the video on learning that no person read it. The honest position is that this is a low-stakes disclosure that costs almost nothing, and the cases where creators regret disclosing it are rare.

Reconstruction. Depicting something that genuinely happened, using generated imagery, in the way documentary has used dramatic reconstruction for decades. The established convention there is to label it visibly as reconstruction, and that convention transfers directly. What does not transfer is doing it without the label because the imagery now looks photographic.

Composite people. A generated presenter who represents no real person. There is no consent question, and there is still a viewer forming an impression that a person is speaking to them. Where the content is instructional this is generally accepted; where it involves recommendation or testimony it is not, which is part of the wider question of where avatars hold up and where they break.

Placement and Wording

A disclosure that exists and is not seen has satisfied a rule and done nothing else.

Put it where the viewer encounters it before forming the impression you are correcting. On screen in the opening seconds, or as the first line of the description — not at the end of a long description, not in a pinned comment, not on a separate page.

Wording is where most disclosures go wrong, in two directions. Vague language — "this video may contain AI-generated elements" — tells the audience nothing and reads as evasion. Over-elaborate legal phrasing signals that someone is worried.

What works is a plain sentence naming what was generated:

The presenter in this video is a synthetic avatar. The script was
written and reviewed by our team.

Narration in this video uses a synthetic version of Dana Okoro's
voice, recorded and used with permission.

The city footage in this section is generated illustration, not
documentary footage.

Each of those states a specific fact and stops. That specificity is what makes a disclosure read as confidence rather than as an apology.

Disclosure Is Not an Apology

A pattern worth resisting: writing the disclosure as though something needs excusing.

Hedged, defensive wording — some elements may have been created with assistance from artificial intelligence tools — communicates uncertainty about whether the thing being disclosed was acceptable. Audiences read that uncertainty accurately.

A plain statement of fact reads differently. It says the creator knew what they were doing, decided it was appropriate, and is telling you. That framing is available at no cost and it is largely a matter of grammar: state what was generated in the active voice, name it specifically, and do not add a justification.

There is a related decision about tone. Where the generated element is central — an avatar presenter, a synthetic narrator — the disclosure can be part of the video's voice rather than a notice bolted onto it. A single line in the opening acknowledging the presenter is synthetic, delivered in the same register as the rest, does the job and stops the question hanging over the remaining three minutes.

Building It Into the Process

Where a real voice is involved, the permission side runs ahead of all of this and has its own scope questions — what the voice may say, for how long, and how consent is withdrawn — set out in consent and disclosure when using a cloned voice.

And where generated footage is being considered for a shot that has to carry factual weight, it is worth checking first whether the technology can even do it well. Many disclosure dilemmas dissolve once you find that the shot would not have worked anyway, which is the territory of what AI video generation can and cannot do right now.

Next Steps

Go through your last three published videos and list every element a viewer might reasonably assume was filmed. If any of them were generated and unlabelled, that is a gap worth closing now rather than after someone else finds it.

Then write your disclosure categories down — one page, which content types get a label and what it says. Having that decided in advance is what stops the question being reopened, badly, on every deadline. The rest of the generation material sits under AI video generation and avatars.