The law can see the difference. It just cannot say it.

What Article 50 requires

The transparency obligations in Article 50 have applied since 2 August 2026. The parts that concern people who publish:

Art. 50(4) — AI-generated text

Deployers who publish AI-generated or AI-manipulated text "with the purpose of informing the public on matters of public interest" must disclose that it was artificially generated. The obligation does not apply where the content "has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication".

Art. 50(4) — deepfakes

Deployers of image, audio or video deepfakes must disclose the artificial generation or manipulation; for evidently artistic, creative, satirical or fictional work the duty is limited to a disclosure that does not hamper the enjoyment of the work. The human-review exemption does not extend here. It is written for text only, so a reviewed, signed-off deepfake still carries the duty — the one place where a higher level on this scale does not relieve you of anything.

Art. 50(2) — machine-readable marking

Providers of generative AI systems — the people who build the model or system, not the people who publish with it — must ensure outputs are marked in a machine-readable format and detectable as artificially generated, so far as technically feasible. The duty falls away where the system performs "an assistive function for standard editing" or does not substantially alter the input or its semantics. Systems already on the market on 2 August 2026 have until 2 December 2026 to meet it.

Art. 50(5) — how

The information must be provided in a clear and distinguishable manner, at the latest at the first interaction or exposure, and must meet the applicable accessibility requirements.

What the Guidelines put outside the scope

The Commission adopted its final Guidelines on Article 50 on 20 July 2026. They read the marking duty narrowly, and several exclusions matter to anyone using this scale:

• Source code, broadly defined — including natural-language comments that form part of it, SDKs, SQL, infrastructure-as-code, YAML and JSON configuration, schemas, scripts, machine-readable specifications, APIs and software libraries.
• Short outputs — single words, captions, alt-text, interface labels.
• Machine-to-machine output processed automatically with no human exposure.
• Intermediate outputs inside a closed workflow; the final output still counts.
• AI-generated translation, now grouped with standard editing alongside grammar correction and spellchecking. Summaries and substantive rewrites still require marking.

Read that list against this scale and one thing follows: the code surface of a declaration is not a legal artifact. Article 50 does not ask you to mark AI-written code, and this scale never claimed the law did. Declaring it is a choice you make for your readers, your collaborators and your own record — which is the only reason any of the six levels was ever worth stating.

The Code of Practice, and the EU icons

The Commission published the Code of Practice on Transparency of AI-generated Content on 10 June 2026, and confirmed it as an adequate voluntary tool for demonstrating compliance in July 2026. Signing is voluntary; Article 50 is not. Section 1 binds providers to marking and detection, Section 2 binds deployers to labelling deepfakes and public-interest text.

With it came a free set of EU icons — SVG and PNG, four contrast variations, no attribution required:

Basic icon

AI was involved. Designed to be supplemented with an interactive second layer that says how.

Fully AI-Generated

The entire content is generated by AI, with no human-created elements.

Partially AI-Modified

Pre-existing, human-made content was partially modified with AI.

The icons must be perceivable at first exposure, embedded so they survive resharing and downloading, and — the Commission's own user testing found this measurably helps — carried alongside a plain-language word such as modified. Using them is optional. The labelling duty is not.

The basic icon is specified as a first layer that expects a second one behind it. That is exactly the shape of a declaration on this scale: a mark anyone can read, and a link to a public definition that pins it down. Put the EU icon where the law wants a mark, and point its second layer at your level page. The icon says AI was involved; the level says what it did.

Where the scale lands on it

The law's test for text is binary — reviewed with editorial responsibility, or not. Those are the two states this scale names as Level 4 and Level 5. The icon column is the nearest EU mark for the same work:

Whether a given page "informs the public on matters of public interest", and who counts as the deployer, are questions for a lawyer, not a badge. The scale does not decide them; it gives you the words for whichever side of the line you are on. The Level 3 row is the one genuinely unsettled mapping, and it is marked as such rather than guessed at.

Outside the EU

Three jurisdictions now ask the same three questions — was AI involved, how much of the work is its own, and did a human take responsibility before publication. They disagree about wording and about who must answer. A declaration on this scale answers all three at once, in terms that survive the translation between them.

China — in force since 1 September 2025

The Measures for Labeling AI-Generated Synthetic Content and the mandatory national standard GB 45438-2025 require both an explicit label a reader can see and an implicit label in the metadata, carrying the nature of the content — confirmed, possible or suspected AI-generated — plus the service provider and a content ID. The two-layer shape is the same one the EU icons assume, and the same one a badge plus ai-usage metadata already produces.

California — operative since 2 August 2026

The California AI Transparency Act (SB 942, as amended by AB 853) puts a manifest disclosure on the surface and a latent disclosure in the asset: provider name, system name and version, a timestamp, and an identifier linking the content back to the system that made it. AB 853 moved the operative date to 2 August 2026 to line up with Article 50; hosting-platform duties follow on 1 January 2027. The duties fall on large providers and platforms, not on individual publishers.

Everywhere else

No duty, and the declaration still works. That is the point of a level that means the same thing in Warsaw, Shenzhen and Sacramento: you are not writing to a regulator, you are telling a reader how the thing in front of them was made.

What to put on an automated page

For a Level 5 pipeline, the declaration is one sentence and two lines of metadata — visible at first exposure, machine-readable, linked to a public definition:

Generated by AI and published without human review. Level 5 on the AI Usage Scale.

<meta name="ai-usage" content="5">
<meta name="ai-usage-standard" content="https://usagescale.org">
<link rel="ai-disclosure" href="https://usagescale.org/5">

The full kit — marks, all three sentence lengths, per-level metadata — is on each level's page. Note that ai-usage metadata is this standard's proposal and the ai-disclosure value is experimental; neither is, by itself, a guarantee of compliance with Article 50's marking or disclosure duties.