Writing content with AI has become so easy that the content written by AI has always been in controversy. One should learn to write accurately from AI, keeping this accuracy in mind there are some things on which the government wants to control. For example, it is mandatory for the government to differentiate between emails or reports written by a human or an AI model, whether a letter or text is written by that person. Based on this, Anthropic has taken a big step. The company recently informed that it is now implementing watermarking on the text generated by its AI assistant Claude, this is directly related to following the new AI transparency law of the European Union.
Table of Contents
Photo : unsplash.com
What is watermarking?
According to Anthropic, watermarking doesn’t mean it will be present everywhere in the text. It will act as a hidden watermark, invisible to the naked eye. There will be no visible markings. The content writer will not even notice the presence of watermarking. Furthermore, it will not be a hidden character added to the text. It will be something like this: a specific pattern is left within the text, which can be read by the person who has the key.
Essentially, when a language model generates text, it creates sentences by selecting words one at a time. Each time, the model has a list of possible words, and normally, it chooses one of these words at random, provided it fits the context. But with watermarking enabled, Claude no longer chooses words completely at random, but rather based on a fixed “key.”
This can be illustrated by an example. Anthropic provided an interesting example. Suppose the key is made up of the digits of pi. If the sequence is 3.1415926535, the first digit, 2, indicates that the sixth word in the list should be selected, then the fifth, then the third, and then the fifth again.
In this way, words are selected each time according to a specific pattern, which can later be used to identify whether the text actually came from Claude. This approach is largely inspired by Google DeepMind’s SynthID-Text technology, which his team described in detail in a research paper published in the journal Nature.
Read More: Why Are RAM Prices So High in 2026?
Will this affect the quality or speed of the response?
So the answer is no. According to Anthropic, this watermarking doesn’t affect the quality of Cloud’s responses or delay response times. It doesn’t require any additional tokens, meaning the process remains the same for users, with a new system behind the scenes.
This technique also has limitations
However, Anthropic himself admits that this method has some limitations. It cannot be determined with certainty whether the text was entirely written by Claude or merely edited. This means that if you use Claude to simply edit a text, it will still be watermarked. This watermark will even be applied to translations. If the text is very short, or Claude has only done some light proofreading, the watermark may not be clearly visible. Interestingly, this watermark is not easily removed by just minor editing; the text will have to be completely rewritten to remove it.
Questions have also been raised about the coding, as a completely AI-generated codebase could pose copyright complications. Anthropic explains that code typically has less watermarking because precise output is required when writing code, and there’s less flexibility in word choice than there is in regular text.
Photos will also be affected.
In addition to text, Anthropic is also adding a cryptographic signature to its images, indicating in the metadata that they were generated by Cloude. The company also stated that this watermarking is currently being applied to Cloud output across all regions, as it is not yet possible to implement it differently for different regions. This change is effective immediately for models released after the beginning of August, and will be added to older Cloud models in the coming months.
Read More:
