AI Usage Scale — releases

1.2.0

Sun, 09 Aug 2026 00:00:00 GMT

The law caught up, so the mapping did too. The scale is unchanged — same six levels, same definitions, same decision tree — so no declaration already published is affected. A minor version, by the policy above.
Every level now carries its EU mapping. levels.json gains euAiActIcon and a note per level, mapping the scale onto the three EU icons published with the Code of Practice: Level 2 is an exact match for *Partially AI-Modified*, Levels 4 and 5 for *Fully AI-Generated*, and Levels 0 and 1 need no icon at all. Level 3 is marked as contested rather than guessed — its substance is human and its form is not, so neither EU term is true of it, and the basic icon plus its interactive second layer is the honest mark.
interop gains three legal regimes beside W3C, IPTC and C2PA: the EU AI Act with its icons and Code of Practice, China's labelling Measures with GB 45438-2025, and the California AI Transparency Act (SB 942 as amended by AB 853). All three ask the same three questions — was AI involved, how much of the work is its own, and did a human take responsibility.
A correction on the EU AI Act page. It read as though human review switched the disclosure duty off generally. It does not: the Article 50(4) exemption is written for text only, so a reviewed, signed-off deepfake still carries the duty. That is now stated where the mistake was.
Source code is outside Article 50. The Commission's final Guidelines (20 July 2026) exclude it broadly — code comments, SDKs, SQL, infrastructure-as-code, YAML and JSON configuration, schemas, scripts, APIs and libraries — along with short outputs and machine-to-machine output. The page says so plainly: the code surface of a declaration is not a legal artifact, and this standard never claimed the law asked for one.
The page is current. Article 50 has applied since 2 August 2026, the Code of Practice was confirmed adequate in July 2026, and the remaining deadlines — 2 December 2026 for marking on systems already on the market, February 2027 for watermark interoperability — are named.
242 corrections across all 21 translated locales. Each locale was audited against English for wording that would send a reader to a *different level*: a dropped negation, "most" softened to "some", "review" collapsed into "edit". Polish had narrowed "the expert" to "the surgeon" and rendered Level 5's "legitimate" as "legal"; Ukrainian, Russian and Turkish carried comparable slips. Key parity and interpolation placeholders are unchanged, and both are gate-tested.
A wider document. Prose keeps a readable measure while tables, marks and snippets now span the full page, so a three-column mapping table is no longer squeezed into a prose column. Section headings open on a rule, level tables gained hairline rows and a header, and a wide table scrolls inside its own box instead of pushing the page sideways. No horizontal overflow at 320px.

1.1.1

Wed, 22 Jul 2026 00:00:00 GMT

Site hardening, accessibility, and documentation. The scale is unchanged — same six levels, same definitions, same decision tree — so no declaration already published is affected. A patch, by the policy in GOVERNANCE.md.
The English-only reference pages now declare English. ecosystem, integrations, cases and eu-ai-act are served under every locale but keep English bodies. They were emitting and hreflang for translations that do not exist; they now declare lang="en", canonicalise to their English URL, and drop the false alternates — fixing a WCAG language mismatch and 84 duplicate URLs.
A strict Content-Security-Policy. The site ships default-src 'self', connect-src 'none', hashed scripts and styles, plus X-Frame-Options and a Permissions-Policy — the "nothing phones home" promise, now proven at the browser level. A build gate fails if the inline theme script's hash ever drifts out of sync.
New docs: a [threat model](docs/THREAT-MODEL.md) — how a self-declaration can be misused, and what the design does and does not defend — and a [FAQ](docs/FAQ.md).

1.1.0

Mon, 20 Jul 2026 00:00:00 GMT

A boundary clarification, two new reference pages, and a larger badge family. No level changed its meaning; under the policy in GOVERNANCE.md this is a minor version.
Boundary clarified: substance the model originated is new material. The decision tree's second question said ideation that leaves no model-produced material in the work is Level 1, while Level 1's own excludes said AI restructuring your argument is not Level 1 — two honest people could reach different numbers. Resolved in the direction the levels always promised ("It invented nothing" has to mean it): a thesis, argument, or structure the model originated is new material that survives into the work, even if you wrote out every word yourself. That is Level 2 or above; feedback, critique, and fact-finding you verify stay Level 1. New edge case: Substance from AI, wording from you. This resolves a contradiction rather than moving settled works between levels — a clarification, not a breaking change. Updated in all 22 languages.
New page: the case corpus (/cases) — real published works classified, with the walk to the number shown and sources linked: The Beatles' "Now and Then" at Level 0, Zarya of the Dawn at 3 with a Surfaces breakdown, The Guardian's GPT-3 essay and CNET Money at 4, the AP's automated earnings, Heliograf and Air Canada's chatbot at 5, and Sports Illustrated's fake bylines as the undeclared anti-case.
New page: EU AI Act (/eu-ai-act) — Article 50 for publishers: the disclosure duty for AI-generated text from 2 August 2026, the human-review exemption, and how it lands on the Level 4 / Level 5 distinction. Explicitly not legal advice; legal review remains on the roadmap.
Spec: "If you doubt a declaration" — what a reader can do with a declaration they don't believe, in all 22 languages. The spec also documents the whole-site declaration at /.well-known/ai-disclosure.json, the discovery path for the Surfaces object.
Marks: auto builds and a named chip. Every mark now ships in a currentColor build, so one inline snippet is legible on any host theme; the level pages' inline snippets use them. A new chip variant spells the level out — "AI USAGE 3 DIRECTED" — for audiences meeting the scale for the first time; the number-only chip stays the universal default. 54 SVGs, up from 24.
Fixed: the hosted-embed snippet declared width="205" for a 122.5 px mark — dimensions are now read from the generated file; the two lighter text greys now clear WCAG AA for small text in both themes; a shared "Find your level" result link (#level-N) now opens on the result instead of question one.
Signing no longer requires GitHub: the sign page links the contact page as an equal path in every language. The author signed signatories.json.
The draft has an end date. The level boundaries stay open for challenge through 31 December 2026; the first stable release, without the -draft suffix, follows once the roadmap's Toward-1.0 gates close. Details in the roadmap.

1.0.0-draft

Mon, 13 Jul 2026 00:00:00 GMT

Initial draft. Not yet stable — the level boundaries are open for challenge.
Six levels, 0–5, decided by five yes/no questions.
Levels named: Human, Assisted, Co-created, Directed, Prompted, Automated.
Optional per-surface declaration (text, image, audio, video, code, data).
Edge cases settled: translation, transcription, decorative assets, derivative works, live generation, non-generative machine learning.
Mappings onto W3C ai-disclosure, IPTC Digital Source Type, and schema.org.
Badge family: 6 levels × chip/stamp × light/dark, type outlined, CC0.
Published in 22 languages, with a switcher. Every translation declares itself as a machine translation on the page, under the scale's own translation rule.
Social cards, schema.org DefinedTermSet (the levels as a controlled vocabulary), reciprocal hreflang, llms.txt and llms-full.txt (the whole standard in one fetch).