What should a disclosure actually say?

What was done, in a few words. That the outline was generated, that a first draft was produced by a tool and edited, or that the images are synthetic tells a reader something they can act on, which a generic label never does.

Do we need to disclose light editing help?

Generally no, and a policy that requires it will be quietly ignored within a month. The line most organisations settle on is whether a machine produced substance rather than polish, which is a judgement worth making once, in advance, rather than repeatedly and under pressure.

Is this a legal requirement?

Sometimes, and the duties differ by jurisdiction and by what the content is. Treat the legal position as a floor rather than the whole question, because the reputational calculus applies everywhere and is usually what decides how an audience reacts.

Glossary

AI disclosure

Telling an audience that what they are reading, watching or hearing was produced with machine help, and deciding how much of that is worth saying.

In plain terms

Saying that a machine helped make something. The hard part is not the decision to say it. It is working out what to say, because there is a long distance between running a draft through a grammar checker and publishing something written wholesale by a tool.

Why it matters

Because the cost of being found out is far higher than the cost of saying so, and because audiences are increasingly good at spotting the difference. An organisation with no settled position ends up deciding case by case, which is how inconsistency and then embarrassment happen.

How it works

The useful question is what to disclose rather than whether. A single label covering every degree of machine involvement tells a reader nothing they can act on, which is why the practice is drifting towards saying what was done rather than announcing that something was.

Degree matters more than presence. Checking grammar, drafting an outline, generating a first version and publishing output largely unedited are four different things, and a policy that treats them identically will either over-label everything or quietly under-label the case that mattered.

Labelling everything trains people to ignore labels. A notice attached to all material carries no information, and the eventual cost is that the disclosure people actually needed to see arrives looking exactly like the ones they learned to skip.

Internal and external disclosure are separate decisions. Telling colleagues how a document was produced affects how carefully they check it; telling customers affects how much they trust it, and the right answer in one direction is often not the right answer in the other.

The asymmetry favours saying so. Disclosing costs a line and a small amount of perceived polish, whereas being discovered costs credibility across everything else you have published, including the parts nobody helped with.

Disclosure is about the reader rather than about the tool, which is the reframe that makes the rest straightforward. The test is whether somebody would feel differently about the work on learning how it was made, and that question can be answered honestly in a few seconds without knowing anything technical.

A settled position beats case-by-case judgement. Deciding in advance which categories of work carry a note, and what that note says, removes the judgement from the moment somebody is under deadline pressure, which is exactly when the wrong call gets made.

How much of this was the machine?

Most disclosure policies fail because they are written as though this were a switch rather than a range. Put the switch at the left and everything carries a label, which makes the label meaningless within weeks. Put it at the right and the cases people would actually want to know about go unmarked. The line that survives contact with real work sits somewhere around the middle, where a machine starts contributing substance rather than polish, and the reason it works is that it matches how an audience reacts: nobody feels misled to learn that a spellchecker was involved, and people do feel misled to learn that the argument they found persuasive was generated. Where exactly to draw it matters less than drawing it once, in advance, and writing down which side each common kind of work falls on. That converts a judgement made under deadline pressure into a lookup, and deadline pressure is reliably where the embarrassing call gets made.

Seen in the wild

A generated deck where the useful note says which parts were drafted and which were checked.
Gamma
A synthetic voice in a recording, identified because an audience would want to know.
ElevenLabs
Marketing copy drafted by a tool and edited heavily, where the honest label is not obvious.
Writer

Common misconceptions

People assume

Adding an AI-assisted label settles it.

In fact

That phrase covers everything from a grammar check to wholesale generation, so it tells a reader nothing about what they are looking at. A disclosure that does not describe the degree of involvement is decoration.

People assume

More disclosure is always better.

In fact

Labelling everything is the same as labelling nothing, because a notice that appears everywhere carries no information. The cost is paid later, when the disclosure that genuinely mattered looks like all the ones people learned to skip.

Questions

What should a disclosure actually say?: What was done, in a few words. That the outline was generated, that a first draft was produced by a tool and edited, or that the images are synthetic tells a reader something they can act on, which a generic label never does.
Do we need to disclose light editing help?: Generally no, and a policy that requires it will be quietly ignored within a month. The line most organisations settle on is whether a machine produced substance rather than polish, which is a judgement worth making once, in advance, rather than repeatedly and under pressure.
Is this a legal requirement?: Sometimes, and the duties differ by jurisdiction and by what the content is. Treat the legal position as a floor rather than the whole question, because the reputational calculus applies everywhere and is usually what decides how an audience reacts.