News
3 min read
Anthropic is building invisible watermarks into texts generated by its new Claude models. This makes it detectable for automated systems that the content was created by AI, even though human readers notice nothing. The measure applies to all Claude products and applications worldwide.
Claude models launched on or after 2 August 2026 will include the watermarks from the outset. Models released earlier will follow at a later stage. The date is not arbitrary: 2 August 2026 is when the core obligations around transparency and high-risk AI systems under the EU AI Act entered into force.
Anthropic applies two techniques depending on the type of output. For text, signals are hidden in the choice of letters and words. For files such as .svg, .png and .jpg, the company uses digitally signed provenance metadata via the open C2PA standard. Both methods are invisible to the reader but legible to automated systems.
The watermarks embedded in running text are not separate tags or hidden characters that can be found in the source code. They are encoded in the statistical choices the model makes when generating words and sentences. A reader sees ordinary text; a detection system recognises the pattern as AI output.
A key characteristic is that the watermarks survive copying and pasting. According to Anthropic, they can also withstand limited human edits. At the same time, the company stresses that the technique is not foolproof. Heavy editing, very short texts or further processing can dilute or erase the watermark entirely. Moreover, a watermark can indicate that AI was involved in creating a text, but not who was responsible for the final content or what changes were made afterwards.
For images and other files, Anthropic uses the C2PA standard, an open specification for provenance metadata that is widely supported in the media sector. Through digital signing, the origin is recorded in the file itself.
The European Union requires providers of AI systems to inform users when they are dealing with AI-generated content. The first provisions of the AI Act came into effect in August 2024, but enforcement of the core obligations around transparency and high-risk applications began on 2 August 2026. Chatbots, deepfakes and AI-generated texts must be recognisable as such.
Companies that breach the rules risk fines of up to 35 million euros or 7 percent of global annual turnover, whichever is higher. For a company of Anthropic's size and valuation, these are not symbolic amounts. Following a funding round in May 2026, the company's valuation was reported to be approximately 965 billion dollars; on secondary markets, shares were being traded at valuations approaching 1.2 trillion dollars.
Other major AI providers face the same obligation. Anthropic's move illustrates how companies are deploying technical measures to comply with European transparency requirements, although the question remains how effective those measures will be in practice when texts are heavily edited or are very short.
The watermarks are applied at model level and are therefore active across all Claude products and applications, regardless of which platform or API is used to generate the output. This makes the approach broadly applicable, but also exposes its limitations.
Once a text passes through multiple hands, is substantially shortened or sections are rewritten, the reliability of detection decreases. Anthropic acknowledges this: the system is not conclusive proof, but an indication. Detecting AI involvement and establishing responsibility are two different questions, and watermarks currently only answer the first.
That said, the approach aligns with what the EU expects from providers: a technical measure that enables systems to identify AI-generated content. How detection companies, platforms and regulators will use these watermarks in practice remains largely an open question.
For European founders and developers using Claude via the API, this move has direct consequences. Texts generated by their products will from now on carry a watermark, even when that output reaches end users. This may be relevant for applications in media, legal services or other sectors where the provenance of texts carries more weight.
At the same time, Anthropic's approach sets a benchmark for how large AI providers handle European transparency obligations. Policymakers and regulators are watching closely: the AI Act prescribes an obligation of result, not a specific technical method. Whether text watermarks are deemed sufficient when they disappear under the pressure of editing will likely become the subject of further scrutiny and possibly legal interpretation.
This article draws in part on the following sources.
AI agents linked to OpenAI exploited a 25-year-old German programming wiki as a communication channel from May to early July 2026, making more than 15,000 edits. OpenAI was aware of the incident weeks before it became public, but did not disclose it itself.
DeepSeek intends to deploy 160,000 Huawei Ascend 950DT chips in a data centre in Ulanqab, Inner Mongolia, dedicated exclusively to inference. Production constraints at Huawei make full delivery before end-2027 or later unlikely.
HumanX Amsterdam opens on 22 September at the RAI with around 200 speakers, five tracks and three days, featuring founders from Legora, Lovable and Celonis alongside enterprise buyers from Diageo, ING and KLM.
