In August 2026, Anthropic announced the implementation of a new watermarking technology for texts generated by its language model, Claude. This decision was a response to growing regulatory demands, particularly within the framework of the European Union's Artificial Intelligence Act. The watermarks will be invisible to users but will allow for the determination of the likelihood of AI involvement in text creation.

How does the watermarking technology work?

The watermarking technology is based on altering insignificant decisions during word selection. The language model already chooses between several semantically similar options, and the watermarking system alters the source of randomness in such a way that a statistically verifiable pattern is formed in the text. This pattern cannot be detected visually, but a specialized tool with the corresponding key can assess the probability that the text was generated by Claude.

It is important to note that the watermarks do not contain identifying information and do not allow one to establish who specifically used Claude, which organization did so, or in which chat the text was created. They will be characteristic of Claude models and are intended primarily for third-party tools to assess the probability of AI involvement in text creation.

Impact on generation quality and cost

According to Anthropic, the technology will not have a noticeable impact on the quality and content of Claude's responses. Text with a "watermark" will not differ visually from ordinary text, and no hidden symbols or other additional data will be added to it. Watermarks will also not require additional tokens, so their use will not lead to an increase in the cost of requests.

For the average user, practically nothing will change: Claude will continue to generate and edit text just as before. It will be impossible to distinguish text with a watermark from ordinary text by eye, and a specialized tool will be required to detect it.

Technology effectiveness in different contexts

The effectiveness of the technology depends on the volume of text. Watermarks are easier to detect in long fragments, whereas in short texts and factual excerpts, there are fewer opportunities for statistical analysis. If Claude is only editing text written by a human, the watermark may practically disappear. It also works less effectively in computer code, as the choice of phrasing there is more strictly limited by syntax and result requirements.

Contradictory data

Some experts express concerns that the implementation of watermarks could lead to new forms of censorship and content control. On the other hand, proponents of the technology argue that it is necessary to combat disinformation and copyright violations. For now, there is no consensus on exactly how this technology will affect the development of AI and its use in society.

Conclusion

The implementation of watermarks in texts generated by Claude has become an important step in the development of artificial intelligence technologies. This solution will allow for better control over AI usage and compliance with regulatory requirements. However, like any new technology, it raises certain concerns and requires further discussion and analysis.