Documentation menuMarking layers

Marking layers

Record which techniques mark your AI content and check them against the two-layer expectation of the Code of Practice.

The Code of Practice on marking and labelling asks providers to apply at least two machine-readable techniques to AI-generated audio, image and video, for example signed metadata plus an imperceptible watermark. For free-form text one technique is enough, and texts under about 150 words (200 tokens) are exempt from watermarking. A visible label is a different duty (Art. 50(4)) and does not count towards the two.

Witness 2.0 lets you write down which layers you apply and checks them against that expectation.

Layer	What it is	How Witness helps
`visible`	A label people can see	`<witness-label>`, `labelHtml`, `<AiContent>`
`metadata`	XMP / IPTC, `data-ai-*` attributes, JSON-LD	`markFile`, `markingAttributes`
`watermark`	A signal in pixels, samples or text	The text watermark. For images and audio use your provider's watermark and add `watermark` yourself
`credentials`	A signed C2PA manifest	Reading: `readC2pa`. Signing: Witness Pro

Record layers on a marking

import { createMarking, markingAttributes } from "@sweberdev/witness";

const marking = createMarking({
  generator: "image-model",
  layers: ["visible", "metadata", "watermark"],
});
markingAttributes(marking);
// { "data-ai-schema": "2", "data-ai-layers": "visible metadata watermark", ... }

Without layers, createMarking records visible and metadata, which is what Witness itself applies.

Check content against the Code

import { assessLayers } from "@sweberdev/witness";

assessLayers("image", ["visible", "metadata"]);
// { required: 2, present: ["metadata"], satisfied: false, missing: ["watermark", "credentials"], notes: [] }

assessLayers(contentType, layers, { words }) takes text, image, audio or video.

Inspect a file

import { inspectLayers } from "@sweberdev/witness";

const { contentType, layers } = inspectLayers(bytes);
// layers: ["metadata", "credentials"]

inspectLayers finds metadata (an AI source type in the XMP) and credentials (an embedded C2PA manifest). It cannot find a watermark: only the provider's detector can. The same on the command line:

npx witness layers ./out/*.png --assume watermark --require

--assume watermark adds the layers you applied but Witness cannot see. With --require the command exits 1 when a file has fewer machine-readable layers than the Code asks for, which suits a CI job.

This is a record and a check of what is present, not a conformity assessment. Whether a technique is robust enough is for you and your legal adviser to judge.