Anthropic recently informed Claude API customers that it will add text watermarks to four existing models starting from November 3, including Claude Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6. This technology has been used in newer models such as Claude Opus 5.5, mainly to identify whether text has been generated or processed by Claude. The watermark will not add additional text or consume more tokens, nor will it affect the output format and billing method of the API. However, two of the models have special limitations. If developers adjust the sampling parameters by themselves, the generated text may not have a watermark.
According to foreign media "The Decoder", Anthropic has sent notices to companies using the above models explaining the new text markup mechanism. The first wave is scheduled to be launched on Anthropic's own Claude Platform on November 3. Customers using the same model through Amazon Bedrock may have to wait a few more days. Since this feature is implemented at the model level, it will be extended to other Claude products and cloud platforms that support this model in the future. Anthropic announced Claude's text watermark technology as early as August 14 this year, and stated that new models launched after August 2 will have machine-readable content tags from launch. This time, the same mechanics are gradually being added to older models. Currently, new models such as Opus 5.5, Sonnet 5.5, and Haiku 5.5 already support text watermarks, while earlier versions such as Opus 4.5 and Sonnet 4.5 are still being introduced in stages.
This kind of watermark is different from the usual logo printed in the corner of the picture, nor is it the insertion of special characters invisible to the naked eye into the article. Anthropic uses a variant of the SynthID-Text technology previously proposed by Google DeepMind. When the model generates text, it leaves patterns with statistical characteristics in the word selection process. When large-scale language models generate text, they often encounter several candidate words with similar meanings that are suitable for inclusion in sentences. Watermark technology adjusts the sampling algorithm between these choices, allowing a large amount of text to accumulate recognizable patterns. Only systems that master the corresponding detection methods and keys can analyze whether this pattern exists. Anthropic claims that internal testing found no significant differences in text quality, creativity, and readability, and the impact on production speed was also quite limited. Because the watermark is embedded in the way the text is selected, rather than information attached to the file, copying Claude-generated articles into email, documents, or other platforms will often still preserve the relevant characteristics. Some minor modifications may not make the watermark disappear, but if it has been significantly rewritten, translated or deleted, the detection effect may be reduced.
There is one exception to this update that API developers should be aware of. According to the notice sent by Anthropic, Claude Opus 4.6 and Sonnet 4.6 will only apply text watermarks when the three sampling parameters of temperature, top_p, and top_k all maintain the default values. If one of the settings is changed to a non-default setting, the watermark will not take effect. These three parameters mainly control the degree of randomness and the selection range of candidate tokens when the model generates content. Many developers will make adjustments based on application needs, such as making the output more stable or increasing the degree of text variation. Therefore, even if the same model is used, different API settings may produce differences in the presence or absence of watermarks. Anthropic specifically reminds that if an enterprise needs to rely on watermarks to meet compliance requirements, it should retain the default settings of relevant parameters.
Anthropic promotes watermarking mainly to comply with the transparency requirements of Article 50 of the European Union’s Artificial Intelligence Act (AI Act). The law will come into effect on August 2, 2026, requiring providers of generative AI systems to, to the extent technically feasible, allow text, images, audio and videos generated by AI to bear machine-readable and detectable marks. For systems that have been launched before August 2, the EU will set up a transition period, and relevant content marking and detection obligations must comply with the regulations before December 2. Anthropic signed the EU's Code of Conduct on Transparency in AI-Generated Content in July and decided to apply the text markup to supported regions globally, rather than just EU users. The company previously explained that at this stage, it has not yet established a marking mechanism that can reliably distinguish between different regions in the long term, so it adopts a globally consistent deployment method.
It is worth noting that EU regulations do not impose exactly the same obligations on model suppliers and content publishers. Model suppliers are required to provide machine-readable markup, and institutions that use AI to publish public interest-related information may additionally be involved in the responsibility of disclosing the extent of AI involvement to readers. According to the current EU regulations, exceptions to the obligation to disclose may apply to text that has undergone substantial manual review, editorial control, and editorial responsibility by relevant personnel; simple correction of grammar or typos does not constitute sufficient manual review.
Another question for publishers, news outlets and corporate content teams is how to check watermarks. Anthropic has provided a dedicated text watermark detection API, but it is still in the private preview stage. It is mainly provided for regulatory agencies, law enforcement agencies, media, fact-checking organizations, researchers, and companies with relevant compliance needs to apply for use. General Claude users cannot directly paste articles into the public tool to obtain official text watermark detection results. Anthropic also reminds that the detection of watermarks can only indicate that Claude may have participated in text generation or processing, and it cannot directly determine whether the entire article was independently written by AI. For example, if the author first completes the manuscript and then sends it to Claude for translation or substantial modification, the output may also leave a watermark. On the other hand, if the mark is not detected, it cannot prove that the article was completely created by a real person.
In addition to text, Anthropic also uses another archival source marking mechanism. When Claude generates files in supported formats such as PNG and JPEG, the system can add a digitally signed content certificate using the C2PA standard to record that the file has been processed by Claude. This type of certificate is stored in the file's metadata and can be checked through the public Claude Content Checker, which is different from the text-embedded watermark. Credentials may also disappear if the file is formatted or metadata is removed.
