AI

Watermarking AI Text: How Content Origin Will Be Verified

Google already watermarks Gemini's text, and Anthropic has announced the same for its future models. How the technology works, what the AI Act requires and what it means for website content.

Watermarking AI Text: How Content Origin Will Be Verified

How can you tell whether a piece of text was written by AI? Until now the answer was "with difficulty". From August 2026 the answer also starts to depend on the model developers themselves. Google already watermarks text generated by Gemini, and Anthropic has announced that text from its future models will carry a watermark too.

How text watermarking works

Anthropic's approach is based on SynthID-text, a technology developed in Google's labs. It's important to understand what it doesn't do: it adds no hidden characters and makes no visible changes to the text.

  1. While generating text, a model constantly chooses between words with similar meanings.
  2. Watermarking subtly skews those choices according to a specific pattern.
  3. Detection is statistical: a tool with a key compares the word choices against the expected pattern and returns a probability.

The result is a probability, not a verdict. Anthropic itself says it can't distinguish "this text was written by Claude" from "this text was heavily edited with Claude".

The regulatory context

According to information cited by Digitalk, the EU AI Act requires new models to add machine-readable marking to generated content, including text, from 2 August 2026, while older models must comply by December 2026. OpenAI, Microsoft, Amazon and Mistral have signed the EU code of practice.

The limitations

  • Watermarking is weaker for code, short texts and strictly factual content, where there is little room in word choice.
  • Paraphrasing with a different model can remove the watermark.
  • Existing detectors such as Turnitin, Copyleaks and GPTZero produce both false positives and false negatives.

Frequently asked questions

Will Google penalise websites with AI-written text?

Google officially assesses how useful and high-quality content is, not how it was created. The problem is mass-producing text with no value. Watermarking doesn't change that principle.

Do I have to label AI content on my website?

That depends on the type of content and the rules that apply. Where images or video could mislead visitors, transparency is sensible practice. For your specific case, consult a lawyer.

What is the best protection for a business?

Original content grounded in real experience, with concrete examples and a clear author. That's why our SEO work focuses on authentic writing, not volume.

Source: Digitalk: Marked. How text will be checked for AI origin (in Bulgarian). This article is an independent analysis by Penkov Studio.

View all

Next step

Got a project in mind? Let us turn it into a website that performs.

Tell us briefly what you need. You will get an email reply with a few clarifying questions and an indicative quote.