What the obligation is
In broad terms, providers of generative AI systems must ensure their outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, with the marking expected to be effective, interoperable, robust and reliable as far as technically feasible.
For text this is harder than for images. There is no file header to write to, since text is copied, retyped and reformatted constantly, so the marking has to live in the words themselves. That is what pushes providers toward statistical watermarking.
This page is an explanation, not legal advice. If a compliance obligation attaches to you, read the text and take advice on it.
What it means in practice for writers
More generated text will carry a mark, which cuts both ways. It makes AI use more discoverable, and it also makes it possible for a writer to demonstrate that a specific document does not carry a specific mark.
It does not make detection universal. Marks remain keyed, open models can be run without marking, and paraphrase degrades the signal. A world with marking obligations is not a world where every accusation can be settled by a test.
How providers are responding
The transparency rules took effect on 2 August 2026. Anthropic said in August that it would watermark Claude text worldwide, and on 5 October OpenAI said it will add a watermark to ChatGPT and Codex text in the EU, using a method it calls textGrain. API developers anywhere can opt in; it is not OpenAI's global default.
Both responses are keyed statistical marks rather than something readers or tools can check openly. OpenAI is limiting detector access to approved researchers and expert organisations. See /guide/chatgpt-watermark and /guide/chatgpt-vs-claude-watermark.
Why a tool sold as reliably defeating the mechanism is a compliance problem
Once marking is a legal transparency mechanism, a tool marketed as reliably stripping it is marketed as defeating that mechanism, which is a different thing from a writer editing their own text. Article 50 obligates the provider of a generative system, not someone editing a document they wrote.
WatermarkRemoverPro draws that line by what it claims, not by refusing to offer editing: the on-device rewrite reduces detectable evidence in your own writing and says exactly that, never "undetectable" and never a promised outcome against a watermark it holds no key for. See /rewrite and docs/REWRITE_PHILOSOPHY.md for the exact claim and its limit.
Wherever this page describes a result: a detected mark is not proof of authorship, and an absent mark is not proof of human authorship. WatermarkRemoverPro's on-device rewrite can reduce detectable evidence but cannot guarantee defeating a vendor's undisclosed watermark, on any tier.