Anthropic has published additional technical details of the invisible marking technology for texts generated by its Claude AI models, in line with transparency requirements set by the European Union's AI Act; a technology that can assess the likelihood that a text was generated or edited by Claude without adding any visible marker.
This technology, known as a "fingerprint" of texts, unlike some common methods, does not rely on inserting symbols, characters, or hidden tokens into the text. Instead, Anthropic subtly alters the model's word-selection process during generation in a controlled way to produce an imperceptible statistical pattern in the output; a pattern that can later be detected with specialized tools.
Anthropic says this method does not add extra tokens to the text and is not expected to have a significant impact on quality, generation speed, or processing cost.
A statistical pattern that reveals Claude's trace
The basis of this technology is situations where Claude faces multiple suitable linguistic options while generating text. In such cases, the system uses a special key and, taking prior words into account, purposefully steers the selection toward certain options.
Repeating this process across a long text leads to the formation of a statistical pattern. A specialized detection tool can examine this pattern and estimate the probability that Claude played a role in producing the text.
Anthropic has stated that the technology used is based on a version of the "SynthID-Text" method developed by Google DeepMind. The company's experiments have also shown that this process does not significantly affect the quality, creativity, or readability of texts.
The longer the text, the more accurate the detection
The performance of this method depends on the nature and length of the text. The longer the text and the more linguistic options the model has to express a concept, the higher the likelihood of detecting the fingerprint.
Conversely, this technology will be more limited for short texts or content where precise word choice is important. For example, large code snippets, where linguistic options are not very diverse, may not be good targets for this method.
This limitation can also apply to texts originally written by humans and then only slightly modified by Claude.
On the other hand, extensive rewriting of a text may weaken or even erase the generated statistical fingerprint. However, translation of a text by Claude can produce a more detectable statistical signature; because in that case the model has greater control over the choice of words in the final text.
Claude's fingerprint does not identify the author
Anthropic emphasizes that this technology is not designed to prove author identity. Detecting the fingerprint can only indicate the probability of Claude's involvement in generating or editing a text and cannot determine which individual wrote it.
The company also said the technology does not disclose information about the user's identity, the organization using it, or the conversation in which the text was produced.
A detection tool is on the way
Anthropic is developing a software interface to detect this fingerprint and says it is developing the technology as part of preparations to comply with the EU transparency requirements.
The company intends to make the tool available globally after launch and gradually extend support for it to older Claude models as well.
Meanwhile, images and files produced by Anthropic's models will use a separate mechanism. This content is intended to be managed based on the C2PA standard and content credentials.
Overall, Anthropic's new approach shows that AI model developers are gradually moving away from visible methods for identifying AI-generated content and toward technologies that enable statistical tracking of origin without noticeable changes to the appearance of the content.