Text

The invisible watermark

watermarkText appends a short payload of Unicode variation selectors to a text. These code points are not rendered, screen readers skip them, and they usually survive copy and paste. The payload carries the AI system, a timestamp and an optional id.

import { readTextWatermark, stripTextWatermark, watermarkText } from "@sweberdev/witness";

const marked = watermarkText(answer, { generator: "Claude", createdAt: new Date().toISOString(), id: "msg_123" });
readTextWatermark(marked); // { generator: "Claude", createdAt: "…", id: "msg_123", raw: "…" }
stripTextWatermark(marked) === answer; // true

Know its limits: it is lost when text is retyped, paraphrased, or passed through a sanitiser that removes unusual characters. Some search indexes and databases normalise text. It is a cheap second layer for chat answers and generated copy, not a robust watermark. Do not watermark content where hidden characters cause trouble, such as code, URLs, IDs or data that is parsed later. stripTextWatermark removes it again.

Emoji also use variation selectors (for example ❤️). Witness only touches its own payload, which starts with a fixed marker.

Page metadata

For an article or page that is mostly AI-generated, add schema.org JSON-LD with digitalSourceType, and optionally meta tags:

import { markingJsonLd, nextMetadata, renderJsonLd } from "@sweberdev/witness";

// any framework
head.insertAdjacentHTML("beforeend", renderJsonLd({ generator: "Claude" }, { "@type": "Article", headline }));

// Next.js App Router
export const metadata = { title: "…", ...nextMetadata({ generator: "Claude" }) };

There is no standard meta tag for AI content yet. Witness writes ai-generated, ai-source-type and ai-generator as a convention; the JSON-LD uses the official schema.org vocabulary.

Visible label

Text published to inform the public on matters of public interest needs a visible disclosure (Art. 50(4)), unless a person reviewed it and someone holds editorial responsibility. The Code of Practice places the label above the text, near the headline. Use labelHtml, <AiContent> or <witness-label> at the top. Mark reviewed content with humanReviewed: true, so the label says so.