Reference
The AI-provenance ecosystem
A declaration is one signal among several. Here is the honest map — what pairs with the scale, and what it deliberately does not lean on.
Cryptographic provenance
Proves which tools and edits touched a file. Answers "what happened to this asset", not "whose thinking is inside it" — the two are complementary, not competing.
- C2PA (Content Credentials) Complements
- An open technical standard for tamper-evident, cryptographically signed content provenance. Its own FAQ states the core spec "does not support attribution of content to individuals or organizations." Provenance can be proven; contribution can only be declared. Both belong on a work.
- Content Authenticity Initiative Complements
- A cross-industry initiative implementing C2PA Content Credentials across cameras, editors and platforms. Free to join. We present the AI Usage level as a contribution/review signal that sits alongside a Content Credential, never as a replacement for it.
Disclosure vocabularies
Machine-readable statements about how a work was made. This is the layer the AI Usage Scale lives in, and every level emits into the established ones.
- IPTC Digital Source Type Interoperates
- An established controlled vocabulary for how media was created (human, AI-generated, composite…). Every level maps to a valid IPTC term. The mapping is lossy in one direction on purpose: IPTC cannot tell Levels 3, 4 and 5 apart, because it has no term for whose substance a work carries or whether a person reviewed it.
- W3C AI Content Disclosure CG Interoperates
- A W3C Community Group incubating syntax for AI content disclosure states. Community Group work is incubation, not a W3C standard. Our emitted `ai-disclosure` metadata is experimental and aligned with this discussion; we contribute use cases and tests, and do not claim endorsement.
- schema.org Interoperates
- The shared vocabulary search engines read for structured data. The six levels are published as a schema.org DefinedTermSet (CC0), so a level is citable by machines as a term, not just a page.
Watermarking
Embeds a signal inside generated pixels or tokens. A useful provenance hint for detecting a specific model’s output — orthogonal to a maker’s honest declaration of contribution.
- Google SynthID Complements
- Imperceptible watermarking for AI-generated image, audio, text and video. Answers "did this specific model generate this?" A declaration answers "what role did a human play, and who is accountable?" A work can carry both.
Consent & licensing signals
Whether a work may be used for training. A different question from disclosure — a declaration describes how a work was made and does not grant or deny any training right.
- Spawning / Do Not Train Adjacent
- Opt-out registries and signals (ai.txt, Do Not Train) letting creators refuse AI training use. Consent, not disclosure. Licences, terms, robots.txt and access controls remain authoritative over training; nothing in this standard overrides them.
- Not By AI Adjacent
- A badge asserting content was made without generative AI. A binary human/AI mark that charges for commercial use. The AI Usage Scale is a gradient, is CC0, and starts at Level 0 for exactly the human-made case — no fee, no permission.
Detection
Tools that guess whether text was machine-written. This standard does not rely on detection, and lists it here with the evidence rather than the sales pitch — so the reader can weigh it honestly.
- AI text detectors (GPTZero, Turnitin, Originality.ai, …) Cautioned
- Classifiers that estimate the probability a passage was AI-generated. In a Stanford study, seven detectors flagged 61% of genuine TOEFL essays by non-native English speakers as AI-generated; 97.8% were flagged by at least one. A standard enforced by detection is a machine for accusing the innocent — which is why this one is enforced by public, falsifiable self-declaration instead.