Lrnon

Lesson 7 of 8 · 8 min read · last verified 2026-08-26

Labelling what you made

In this lesson you will:

  • Decide when generated media requires disclosure
  • Explain what provenance metadata can and cannot establish

You have made something. Does it need to say so?

“Always” is the easy answer and it is not right — nobody labels a stock illustration on a slide. “Never” is obviously wrong. The line is more useful than either.

The test

Would a reasonable person take this as a record of something that actually happened?

If yes, say what it is. If no, you are usually fine.

A generated illustration on a blog post: no label needed. It is decoration, and nobody thinks it is a photograph of a real event.

A photorealistic image of a named place, a real person, or a specific occurrence: label it. That is the case where the medium’s implicit promise — this happened, and someone was there — is being borrowed.

This is why L5 flagged video specifically. Video carries that promise far more strongly than a still.

Four places the answer is always yes

Regardless of the test:

  • News, journalism, documentary. Non-negotiable.
  • Anything involving a real identifiable person, including their voice.
  • Evidence — insurance, legal, workplace complaints, any record intended to establish what occurred.
  • Where a rule already applies. Advertising standards, platform policies, your employer, and in a growing number of jurisdictions, the law. The EU AI Act’s transparency obligations are the most cited example; more are arriving.

Provenance metadata, honestly

There is a standard, C2PA, sometimes branded Content Credentials. It attaches signed information to a file — what made it, when, what happened to it since — and support is growing across cameras, editors and generation tools.

It is worth using where your tools offer it. Two things it does not do, and both are commonly misunderstood.

It cannot prove an untagged file is authentic. Absence of provenance is not evidence of anything, because metadata is stripped constantly by ordinary activity: uploading to most platforms, taking a screenshot, resizing, sending through a messaging app. Most honest media in circulation has no credentials attached.

It is not a detector. It travels with files that carry it. It says nothing about the ones that do not.

So treat it as a way of signing your own work, not as a way of checking other people’s. E7 covers verification properly.

Label it visibly, not just in metadata

Following from that: if disclosure matters, put it where a person will see it — a caption, a line in the credits, text in the corner.

Metadata that survives your export and dies on upload has disclosed nothing to anyone. A visible caption survives a screenshot.

The cost is almost nothing

The strongest argument for disclosing is not ethical, it is practical.

The cost of labelling is a line of text. The cost of not labelling, and being found out, is that everything else you have published becomes questionable — including the parts that were real. That asymmetry is not close, and it does not require you to have settled any philosophical question.

Try it now (5 minutes)

Take three pieces of generated media you have made or seen. Apply the test to each.

For any that need a label, write the actual caption — the sentence you would use. It is shorter than you expect.

Check your understanding

1. The test for whether generated media needs a label:
2. Provenance metadata such as C2PA can:
3. Why put disclosure in a visible caption rather than only in metadata?

Recap

Label anything a reasonable person would take as a record of what happened, and always for news, real people, evidence, or where a rule applies. Use provenance metadata to sign your own work, knowing it cannot authenticate anyone else’s, and put the disclosure somewhere a person will actually see it.

🗂 3 flashcards from this lesson join your daily review.

Previous: Voice and the consent line · Next: The line you do not cross