How does magic watermark on AI texts work? OpenAI joins Claude
Texts written via ChatGPT will receive an invisible watermark. The company follows in the same footsteps as Anthropic, which has marked Claude’s lyrics since August. We explain how these types of solutions work.
OpenAI announced that in the coming weeks it will begin marking text produced by ChatGPT and Codex, its programming agent. The change will apply to users in the EU. Could this mean the beginning of the end of AI-based writing? It’s hard to say.
Today, many texts are created in this way in schools and universities, in promotional and marketing materials and e-mails, but also in online information services. Will the markings eliminate this practice?
The chance is small, especially since the magic code for AI text detection, i.e. for verifying whether a given text was created using the LLM (Large Language Model) or not, is only available to companies that create LLMs.
And although companies reveal what their technologies are, they do not indicate how these “watermarks” can be found in texts.
It sounds like magic, a sign hidden in the words
OpenAI’s technology is called textGrain. Does not add visible characters or hidden symbols to text. The model slightly changes the way words are selected, which creates a characteristic statistical pattern in the text. – The character lives in the words themselves and moves with the text when it is copied and pasted – explains OpenAI.
To detect it, you need a program known to OpenAI and a secret key.
Claude was the first, but probably everyone will mark
Anthropic has been tagging Claude’s lyrics since August 2, 2026 and doing so worldwide. The mechanism works similarly. When the model chooses between equally good words, it follows the pattern set by the secret key.
This does not mean that Claude starts choosing strange words to mark up text or deliberately lowers the quality of the text in order to mark up text. The mechanism works at the level of decisions where there are several equally good options. The model chooses one of them, which is the secret key.
Anthropica’s internal testing reportedly showed no practical impact of their system on text quality, creativity, or readability.
To an ordinary reader, the company’s decisions look completely normal and random – at the level of a single material. Only the analysis of a sufficiently long text allows us to check whether the arrangement of these decisions statistically corresponds to the pattern characteristic of the secret key.
According to some sources, Anthropic uses the SynthID-Text technology developed by Google DeepMind. New models have the watermark for now, and older models will receive it by December 2.
Both companies make the change for the same reason. From August 2, 2026, the EU AI Act requires that content created by AI be marked in a machine-readable way.
The sign may be weakened or removed by editing the text
Companies admit that watermarking is not foolproof. In OpenAI tests, a 10% replacement. words into synonyms reduced the detection efficiency from 92 to 66 percent. Short texts, mathematical calculations and translations are also more difficult to detect. Anthropic also acknowledges that redrafting text may dilute or remove the mark.
Both technologies do not answer the question of how much human involvement was in the creation of a given material, nor do they indicate the authors of the text or the owners of the rights. The presence of a sign only means that the text has passed through a given AI model.
Sources: TechCrunch, Telepolis, ITHardware, Business Insider
