Trust and Transparency

Provenance Disclosure

Mark which parts of a document were generated by AI, edited by a person or written by a person.

Phase: Design · Build
Applies to: Developers, end users
Oversight: Human in the loop
Key factor: System autonomy

Overview When to use Where the rules aren't clear yet Questions to ask Future exploration How other systems do it Anti-patterns to pay attention to References Related

Overview

Once AI text is pasted into a document, it's indistinguishable from the rest, so reviewers can't tell what a person checked. Provenance Disclosure shows the origin of each block (AI, AI edited by a person, or written by a person) and carries machine-readable credentials, such as C2PA content credentials, when the content leaves the tool.

When to use

Use when

AI generates content inside documents, reports or media that people edit, approve or publish.

Where the rules aren't clear yet

Unresolved

Minor assistive edits, like spelling or formatting, are excluded: the EU AI Act marking duty doesn't cover assistive editing that doesn't substantially change the input. But no legal or descriptive threshold exists for how much human editing changes the label. Is it enough if a user deletes an em dash, or rewrites the sentence? The system has no rule to decide when AI-generated content no longer needs labelling.

Questions to ask

Ask these at kickoff, when design work starts. Tick them off as your team answers.

Where in the system is AI generating content: documents, reports or media?
Does exported content need machine-readable marking?
Is your company liable for third-party AI-generated images that users import into your system?

Future exploration

Can the tool track origin per block or input, and mark which AI output a person has reviewed?

How other systems do it

One publishing platform asks creators to declare AI use in detail, and shows how the editing question can be handled in practice.

Amazon Kindle Direct Publishing

When publishing a book, authors answer separately for texts, images and translations. For each, they say how much was AI-generated (some sections or the entire work; one, a few or many images) and whether it was edited minimally or extensively, and they name the tools used. Amazon treats content as AI-generated even when it was extensively edited afterwards. The declaration is internal to Amazon: it doesn't appear on the book's product page, so readers never see it.

Amazon KDP form asking whether AI tools were used for texts, images or translations, with text options from None to Entire work with extensive editing — Texts: how much was AI-generated, and how heavily it was edited

Amazon KDP form with the images dropdown open, from None to Many AI-generated images with extensive editing, and a field naming the AI tool used — Images: how many, how heavily edited, and which tool was used

What this shows: a structured answer to “how much editing?” that the creator chooses themselves, so the system doesn't have to judge it. It also shows the gap: the declaration stops before it reaches the reader.

Anti-patterns to pay attention to

AI-generated content loses its origin when you edit or paste it.
Reviewers can't see which passages are AI-generated and unreviewed.

References

Current guidelines · EU AI Act2 references

For orientation only, not legal advice. These obligations apply beyond high-risk systems. EU AI Act links open a readable version of the official text published on EUR-Lex; ISO/IEC 42001 is a paid standard.

EU AI Act Art. 50(2) Providers of systems that generate synthetic audio, image, video or text must mark outputs in a machine-readable format so they're detectable as AI-generated, with an exception for assistive editing.
EU AI Act Art. 50(4) Deployers must disclose deepfakes, and AI-generated text published to inform the public on matters of public interest, unless it has had human review and editorial responsibility.