How it is defined, and how it talks to everything else.

Decision procedure

Five questions, asked in order. The first one that terminates gives the level.

Surfaces

Optional. A single number describes the work as a whole. When parts of it differ sharply, declare them. The badge always shows one number; the breakdown lives behind the link.

Headline rule. The headline level describes the part of the work that carries its meaning. Decorative material that a reasonable audience would not consider material to the work does not drive the headline — declare it here instead.

Recognised surfaces: textimageaudiovideocodedata

{
  "aiUsageScale": "1.0",
  "level": 3,
  "surfaces": { "text": 3, "image": 5, "audio": 0 }
}

Edge cases

Every one of these came from trying to break the decision tree. They are settled here so that two honest people reach the same number.

Translation

A faithful translation inherits the level of the source work and adds a translation note. Translating a Level 1 article with a model publishes it as Level 1, translated by AI.

The substance and structure are unchanged. Penalising translation would make the world's writing less available, which is the opposite of the point.

Transcription

Transcribing speech you recorded is mechanical processing. It does not raise the level.

The words are yours. The model only wrote them down.

Decorative assets

A generated header image on a human-written article does not make the article Level 4. Declare the image under Surfaces.

The headline level describes what carries the meaning. Otherwise one stock image would swallow the whole declaration, and people would simply stop declaring.

Derivative works

Declare the level of your own contribution, and cite the level of the source if it declared one.

Live generation

Content generated on demand and shown to a user without review is Level 5, even if the system prompt was written carefully by a person.

Non-generative machine learning

Autofocus, denoise, upscaling, colour matching, and classification that invent nothing are not generative AI. They do not move you off Level 0.

Interoperability

This scale complements provenance standards. It maps to the established IPTC vocabulary and experiments with the candidate syntax discussed by the W3C AI Content Disclosure Community Group; that syntax is not a W3C standard.

Level mapping

The mapping is lossy in exactly one direction, and the loss is the point: IPTC cannot tell Levels 3, 4 and 5 apart. It has no term for whose substance this is, and none for whether a person read it. Those are the two questions readers actually care about, and they are the two this scale exists to answer.

Machine-readable form

The canonical English definitions live in one file and drive the badges and machine-readable output. Localised explanatory copy is maintained separately and checked for structural parity.

GET /levels.json · CC0

<!-- AI Usage Scale proposal; custom metadata, not a registered standard -->
<meta name="ai-usage" content="3">
<meta name="ai-usage-standard" content="https://usagescale.org">

<!-- Experimental W3C Community Group alignment; not a W3C standard -->
<meta name="ai-disclosure" content="ai-generated">
<link rel="ai-disclosure" href="https://usagescale.org/3">

Prior art

Nothing here is unprecedented, and pretending otherwise would be its own small dishonesty.