Marcus Andrey Study
Ler em português Library

Photography and Imagery · Step 4 · 5 min read

Labels and provenance

How to tell people an image was generated, and how an image can carry its own history. Label wording, what labels cost, C2PA and Content Credentials, invisible watermarks, and what the EU AI Act asks since August 2026.

If people can't tell by looking, you have to tell them. There are two ways: a label people can see, and a record the file carries with it. Both are young, both are imperfect, and both are becoming the norm. A designer decides how the label looks, where it sits and what it says, and those choices change how people read the image.

What you'll be able to do after reading:

  • Choose label wording that people understand and don't over-read
  • Place a label where it does its job
  • Explain C2PA, Content Credentials and invisible watermarks in plain words
  • Know what the EU AI Act asks of generated images

1. Where the label goes

AI-generated loaf of bread, with a label on the image.AI-generatedOn the image: seen first
AI-generated loaf of bread, labelled in the line below it.Image generated with AI
AI-generated loaf of bread, with a small info label in the corner.AI infoIn a detail: quiet, easy to miss
Three ways to label the same generated image: on the image, in its caption, or as a small info marker. The first is the hardest to miss; the last is the easiest to ignore. The middle label is the caption itself.Generated image: made for this course with FLUX 3 Image Klein on a laptop.
  • On the image is the hardest to miss, and it survives screenshots of the image itself.
  • In the caption is calm and readable, but separates from the image when it's shared.
  • In a small info marker is the quietest, and the easiest to miss. Social platforms often use this pattern; Meta's is called "AI info" (Meta, labelling AI-generated content, 2024).

For product imagery where a reader could mistake the image for a record, put the label on the image or right under it.


2. What the words do

Labels are read, and words carry meaning beyond the literal:

  • The wording matters. In a study across the US, Mexico, Brazil, India and China, labels like "AI generated" and "Generated with an AI tool" were understood as saying how something was made. "Deepfake" and "Manipulated" were read as saying it was misleading (Epstein et al., 2023). Say how it was made; don't accuse your own image.
  • Labels have a cost. In experiments with nearly 5,000 people in the US and UK, an "AI-generated" label lowered how accurate people found content and how willing they were to share it, even when the content was true and made by people (Altay and Gilardi, PNAS Nexus, 2024). That study used headlines, not images, but the lesson travels: label what's generated, and only that. A label on everything means nothing.

3. Provenance: an image that carries its history

The imageWhat its Content Credentials can say
A photographCaptured by a camera that signs its files → edited in an app that adds its own signed step → published. Each step is recorded and signed.
A generated imageGenerated by a named AI tool → published. The record says it was made, not captured.
An image with no credentialsNothing. That doesn't mean it's fake: most images online carry no credentials yet.
A credentialed image, re-savedOften nothing: a screenshot or a re-export can strip the record. That's why some tools also add an invisible watermark.
What provenance can and can't tell you, in the C2PA standard used by Content Credentials. A missing record is a question, not an answer.

C2PA is an open standard, from the Coalition for Content Provenance and Authenticity, for attaching a signed record to a media file: what made it, what changed it, who published it (C2PA). Content Credentials is the name you'll see in products that use it, often as a small "CR" mark; anyone can inspect a file's record in the public verify tool (Content Credentials, verify).

Two limits to remember:

  • Records can be lost. A screenshot or a careless re-export can strip them. A missing record means "unknown", not "fake".
  • Watermarks help. Some generators also hide an invisible watermark in the pixels, which survives where metadata doesn't. Google's SynthID is one, with a public detector (Google DeepMind, SynthID).

For your own work: keep the credentials your tools add, don't strip metadata on export, and keep your own record of tool, date and prompt.


4. What the law asks (EU)

Since 2 August 2026, the EU AI Act's transparency rules apply. In short, providers of generators must mark generated content in a machine-readable way, and people who publish realistic generated or manipulated content, "deep fakes", must disclose it; for clearly artistic or fictional work a lighter disclosure is enough (EU AI Act, Article 50).

For a product team in the EU, the practical version: label realistic generated images, keep their machine-readable marks, and check with your legal team for your case. Platforms have their own rules too; YouTube, for example, asks creators to disclose realistic synthetic content (YouTube Help).


5. Common mistakes

  • No label on a realistic generated image, because "it's only decoration".
  • Accusing words like "fake" or "manipulated" for honest generated work.
  • A label so small or so far away that nobody connects it to the image.
  • Stripping metadata on export and losing the provenance record.
  • Labelling everything "AI" just in case, so the label stops meaning anything.

6. Vocabulary

Term Meaning
Disclosure label Visible text saying an image was made or changed with AI
Provenance The history of a file: what made it, what changed it, who published it
C2PA An open standard for signed provenance records in media files
Content Credentials The user-facing name and mark for C2PA records
Invisible watermark A signal hidden in the pixels that tools can detect

7. Try it (no tools needed)

  1. Find three generated images online: on social media, in an advert, on a news site.
  2. For each, note where the label is, what it says, and how long it took you to notice.
  3. Rewrite the worst label so it says how the image was made, in four words or fewer.

Reflection questions: Which label placement would you use in your product, and why? What would you do with an image whose record was missing?


8. If you already practise

  • Design the label once, as a component in your design system, with its wording, placement and size.
  • Check your export pipeline: does it keep Content Credentials, or strip them?
  • Agree with your team what counts as "realistic" and needs a label, with examples, before the first argument.

Checklist

Learn more

What changes when the image is generated · Image libraries and AI · Copyright and AI-generated assets · Licences without fear · Generated image policy template

Saved only on this device.

↑ ↓ to move · Enter to open · Esc to close