Scheme version 1

The scheme, for machines

Everything a publisher, a CMS or a crawler needs to read and write an aiflags grade: what the five grades mean, the attributes to emit, and where a grade sits relative to real in-file provenance.

What v1 promises

Terms published under /schemas/v1/ are frozen. If the meaning of a grade ever changes, that becomes /schemas/v2/ rather than an edit here — otherwise a declaration made today would quietly start saying something else. The same reasoning is why the badge snippets point at a pinned /badge/v1/ URL, and why every declaration carries its scheme version.

The five grades

Value	Name	Means
A	Human-authored	Made by people from start to finish
B	AI-assisted	Human-made, with limited AI-generated elements
C	Co-created	AI-drafted, substantively reworked by a person
D	AI-generated	Predominantly AI-generated, limited human changes
E	Fully generated	Machine-produced; a person only prompted, selected or published

The grade describes production state, and only that. It is not a legal risk class, it does not select an EU disclosure icon, and it does not establish whether Article 50 requires disclosure.

Two layers, two different duties

AI Act · binding

A visible badge is not the machine-readable marking Article 50(2) requires. That duty falls on providers of generative AI systems, and is met inside the file — with standards like C2PA and watermarking — not by markup a publisher adds afterwards.

It does not run the other way either. A deployer cannot rely on the provider’s embedded marking to meet their own disclosure duty: Article 50 requires something a person can perceive without any special tools, by first exposure. In-file provenance and a visible label answer different obligations, and neither substitutes for the other.

Our grade is a third thing again: voluntary, and about authorship rather than disclosure.

Attributes

data-ai-label: The grade: A, B, C, D or E.
data-ai-scheme: Which version of the scale the grade was assigned under — currently aiflags-v1.
data-ai-declared: ISO date the declaration was made. Optional, and useful once a page outlives its author's memory.
data-ai-source: Provenance axis: was anything authentic captured. One of the five source ids below.
data-ai-mode: What the model did: gen-create, gen-edit, style-transfer or enhance-only.
data-ai-tool: The system that did it, in plain words. Optional.
data-ai-disclosure: Which EU icon fits the source axis, if one is being shown: generated, modified, basic or none.

<a href="https://www.aiflags.eu/label/C" rel="ai-label"
   data-ai-label="C" data-ai-scheme="aiflags-v1" data-ai-declared="2026-08-04">
  AI involvement: C — Co-created
</a>

JSON-LD and C2PA

The JSON-LD context defines the aiflags: terms against this page. The C2PA file is an example manifest showing where a grade belongs alongside standard provenance — including an Article 50 assertion that is our proposal in our own namespace, since the Commission has published no official assertion schema for this.

transparency.jsonld c2pa-example.json

Source axis and IPTC

The source axis maps onto IPTC’s registered digital source types — the same vocabulary C2PA writes into files — and is what decides which EU icon fits, if you are showing one.

Source	IPTC digital source type	Fitting icon
pure-synthetic	http://cv.iptc.org/newscodes/digitalsourcetype/trainedAlgorithmicMedia	generated
composite	http://cv.iptc.org/newscodes/digitalsourcetype/compositeSynthetic	modified
authentic-modified	http://cv.iptc.org/newscodes/digitalsourcetype/compositeWithTrainedAlgorithmicMedia	modified
authentic-assisted	http://cv.iptc.org/newscodes/digitalsourcetype/algorithmicallyEnhanced	basic
authentic	http://cv.iptc.org/newscodes/digitalsourcetype/digitalCapture	none

What this scheme does not claim

The label finder gives orientation, not a legal determination.
A grade never selects an EU icon, and never encodes a legal duty.
The EU icons are optional Code of Practice tools, not a legal requirement — and showing one does not by itself establish compliance.
The Code of Practice placements are commitments for Code signatories, not mandatory rules for everyone.
Grade E describes production state only. It does not mean deepfake.
The risk-tier examples are simplified. The Act's precise definitions, scope provisions and exceptions decide real cases.

aiflags.eu is an independent project, not affiliated with the European Union. This page is informational, not legal advice.