Invisible Watermarks in AI-Generated Text

AI-generated text now ships with an invisible watermark embedded inside it. It survives copy and paste. And, after turning it over for a few days, I think it is the biggest shift for content SEO this year. Anthropic published it this week, on 11 August 2026, when it signed the European AI Act's Transparency Code of Practice, whose obligations took effect on 2 August.

How does the watermark in AI-generated text work?

It is imperceptible: it does not change the meaning, the quality or the readability.
It lives inside the text, not in the metadata. That is why it travels with you when you copy and paste.
It is applied at model level: it does not matter whether the text comes from the API, Claude, Claude Code or Cowork.
It applies worldwide, not only in the EU.

Anthropic marks the text, but has no public detector yet

Here is what struck me most when looking at the full picture, and it is the asymmetry that makes the map counterintuitive: Anthropic marks text, but does not have a public detector yet. OpenAI does not mark text yet, but already has a public detector and an API: for now it marks images (C2PA plus SynthID) and audio. One marks and cannot prove it. The other can prove it and does not mark yet. That window closes soon, because OpenAI says its goal is to reach every modality, text included.

And these are not two companies fighting their own battle. The code was signed by 190 organisations: Google, Meta, Microsoft, Mistral, Cohere and Getty Images among them. In the publishers section you find Iberdrola, Bulgari, Lufthansa, Lenovo and the European Court of Auditors.

You can check it today with your own AI-made images

If you want to see it in two minutes instead of waiting for text detectors, take an image from your blog made with AI and upload it to openai.com/verify. It comes back marked. What matters is not that you can do it: it is that anyone can, including your client or the outlet where you place content.

Does AI-written text get penalized? Three caveats before anyone panics

Google does not penalize content for being AI-generated. Its stance has not changed since 2023. It goes after scaled content abuse, not the origin of the text.
The mark says "processed by", not "written by". If you use AI only to proofread or translate your own text, it gets marked all the same.
It degrades. With paraphrasing, translation, heavy editing or very short passages, the signal is lost. So it does not stop whoever runs mass spam: it marks whoever uses it honestly.

My take on watermarks in AI-generated text and your content strategy

That third caveat is the one I find most uncomfortable, because it inverts the incentive: whoever generates a thousand pages and runs them through a paraphraser comes out clean, while whoever writes with judgment and uses AI to polish ends up marked. That is why it does not look like a race worth running. What protects your brand is not that your content carries no mark. It is that it holds proprietary data and judgment nobody else can publish.