Writing a disclosure somebody can use

Two kinds of label

A compliance label exists to satisfy a rule. "This content may contain AI-generated material." It is placed where a rule requires, worded to cover everything, and tells a reader nothing.

An informative disclosure tells somebody what they are looking at. "The presenter's voice is synthetic; the footage is real. The background of the third image was extended using generative fill."

Only the second changes what a viewer understands, and it costs about the same to write.

What makes a disclosure useful

Specific about what. "AI was used" covers a spectrum from spellchecking to fabricating the entire subject. Say which element.

Specific about what was not. Frequently the more valuable half. In factual work, "the photographs are unaltered; the diagrams are generated" tells a reader exactly how much weight to put on each.

Placed where the content is. A disclosure in the terms of service is not a disclosure. It belongs adjacent to the thing, visible at the moment of viewing. For video, in the video or in a persistent overlay, because descriptions do not travel when a clip is reshared.

Written for the reader. Not "generative adversarial synthesis was employed in post-production" but "this voice is a computer, not the presenter".

Proportionate. A stock background image on a corporate page does not need a paragraph. A synthetic voice in a news report does. The test is whether a reasonable person would change how they interpret the piece if they knew.

A usable rule

The threshold worth adopting, because it is both defensible and easy to apply:

Disclose when the synthetic element affects how the audience should interpret what they are seeing.

Under this rule:

• Disclose: a synthetic voice presented as a person; a photograph of an event that did not occur; a person depicted doing something they did not do; a testimonial from someone who does not exist; sound design in a documentary presented as recorded on location.
• Usually do not need to: removing a distracting object from a background; upscaling; colour correction; generated texture in an obviously illustrative graphic; a stock-style illustration used decoratively.
• Judgement: an illustration of a real event; a composite of several real photographs; generated background music in a factual piece.

Notice this is the standard journalism already used for photographic manipulation. Dodging and burning was always acceptable; adding or removing a person was not. The line is whether the change alters the meaning.

Where over-labelling does harm

There is a failure mode in the other direction worth naming, because good-faith organisations are falling into it.

Labelling everything trains the audience to ignore the label. If a "may contain AI" notice appears on every page, it carries no information and the genuinely important disclosure is invisible inside the noise.

Some platforms' automatic labelling has already produced this: photographs that were merely opened in an editor with generative features have been labelled as AI, which is technically defensible and misleads the reader about what happened. The photographers affected reasonably objected, because the label said something untrue about their work.

Accuracy in both directions is the standard. A label that overstates is a false statement about your own work.

The practical shape

For most working people this comes down to two habits.

Keep the record — model, prompt, what was generated, what was captured, what was done by hand. The layered file and the note from earlier modules.

Write two sentences at delivery. What is synthetic, what is not. Put them where they will be seen.

That is the whole practice. It takes a minute, it is honest, it becomes an ordinary professional habit rather than an admission, and it is far cheaper than the conversation that happens when a client discovers it later.

The tone question

A disclosure is often written apologetically, and it does not need to be. There is a difference between confessing and informing. "Some elements of this were created with AI" reads as a hedge; "the landscape is a photograph, the creature is generated" reads as a caption.

The second is better writing and it is also better positioning. Work that states plainly what it is invites judgement on what it is, which is the position you want to be in. Work that hedges invites the reader to wonder what else is being hedged. Museums, science publishers and good documentary producers have handled this for a long time with reconstructions and illustrations, and their convention is worth copying: name the thing, say how it was made, and move on.