Anthropic has provided clarity on its strategy for embedding invisible watermarks into text generated by its Claude AI model, a move aimed at complying with Europe's stringent AI transparency regulations. The company announced on Friday that Claude's text marking system utilizes "a version of the SynthID-Text approach"—an open-source watermarking technology pioneered by Google DeepMind. This innovative system works by generating detectable patterns through the manipulation of wording probabilities.
This watermarking capability, alongside support for C2PA standards for images processed by Claude, is being implemented to fulfill Anthropic's obligations under the European Union's AI Act. This landmark legislation mandates that all synthetic audio, images, video, and text must incorporate machine-readable identifiers, allowing them to be recognized as artificially generated or manipulated. Anthropic assures users that these text watermarks will not lead to increased costs for Claude or "have any practical impact on the quality or content of Claude’s outputs." Here is Anthropic's explanation of the underlying mechanism:
Consider the sentence fragment “The weather today was cold and…”. The subsequent word is highly unlikely to be “sugary,” but words like “overcast” or “grey” are quite probable. In most scenarios, the specific choice between these latter two words does not significantly alter the reader's understanding, as the sentence's core meaning remains largely consistent. Typically, such word choices are determined by a random number generator.
Watermarking capitalizes on these low-stakes choices—which occur numerous times throughout any piece of generated text—to subtly embed a pattern within Claude’s responses. This pattern is imperceptible to the human reader but becomes detectable to anyone possessing the specific decoding key. When watermarking is active, word choices are still made with an element of randomness, but the source of that randomness is altered. Instead of relying on an arbitrary random number generator to select the next word, the watermarking system employs the key in conjunction with the preceding few words to guide the model's selection.
As Anthropic highlights, the EU’s AI transparency directives extend their reach to other major AI developers, indicating that Claude will not be the sole model to introduce text watermarks. Google’s Gemini chatbot, for instance, has incorporated the SynthID Text solution since 2024. While OpenAI has not yet detailed specific text watermarking plans for ChatGPT within its AI Act compliance roadmap, it too will be subject to the requirements of the new law.
The Editorial Staff at AIChief is a team of professional content writers with extensive experience in AI and marketing. Founded in 2025, AIChief has quickly grown into the largest free AI resource hub in the industry.