Anthropic Enhances Text Identification for Claude with Invisible Watermarking
Anthropic has implemented a new system to facilitate the identification of content produced by its artificial intelligence models. Since August 2, 2026, new versions of Claude can integrate an invisible watermark directly into the generated texts, an evolution aimed at meeting the transparency requirements of the European regulation on artificial intelligence.
Unlike traditional marking, the process does not visibly alter the responses provided to users. The watermark is embedded at the very moment the text is generated and is intended to allow, through a detection system, an estimation of whether content was likely produced by Claude.
An Integrated Marking at Generation
The technology chosen by Anthropic is based on SynthID-Text, developed by Google DeepMind. The principle involves intervening in the word selection process: the model follows a predetermined sequence based on a secret key, allowing the presence of the marker in the text to be searched for.
Anthropic presents this approach as a traceability mechanism that does not require adding a visible mention or symbol to the response. The company also claims that this method does not degrade the quality of the generated texts.
Google DeepMind had previously experimented with SynthID-Text on about 20 million responses produced by Gemini. According to the results reported by the company, this experimentation did not reveal a measurable decrease in user satisfaction.
A Detection Tool, Not Identification
However, the watermark should not be considered a means of tracing the author or user behind a piece of content.
A positive detection could indicate that a text exhibits characteristics consistent with generation by Claude, but the system would not identify the account or person who used the model.
This distinction is important in a context where tools for detecting AI-generated content are increasingly used in education, media, and business.
Limitations with Transformed Texts
The system is also not infallible. Very short texts can make detection more difficult. The same goes for hybrid content, combining human writing and automatic generation, or for texts that have undergone significant rewriting.
A new transformation carried out with another artificial intelligence model can also modify the content sufficiently to affect the ability to retrieve the watermark.
This approach thus differs from traditional AI detectors, which generally seek to identify statistical characteristics that might distinguish human production from automatically generated text. However, these tools can still lead to classification errors.
Anthropic is Preparing a Verification Tool
Anthropic also plans to provide a dedicated API for watermark verification. This interface should allow services or developers to search for the presence of the marker in submitted content for analysis.
However, the company has not yet indicated when this API will be accessible.
With this system, Anthropic aims to introduce an additional level of transparency around the content produced by Claude, while maintaining an invisible marking for the user. A technology that could become increasingly important as transparency obligations regarding AI-generated content strengthen.
-
22:00
-
21:41
-
21:21
-
21:05
-
20:45
-
20:30
-
20:15
-
19:47
-
19:32
-
19:15
-
19:00
-
18:42
-
18:25
-
18:15
-
18:10
-
17:47
-
17:32
-
17:15
-
16:50
-
16:32
-
16:16
-
15:56
-
15:40
-
15:20
-
15:05
-
14:45
-
14:30
-
14:15
-
14:00
-
13:42
-
13:25
-
13:10
-
12:47
-
12:31
-
12:15
-
12:00
-
11:42
-
11:25
-
11:11
-
10:47
-
10:32
-
10:15
-
10:09
-
10:00
-
09:45
-
09:30
-
09:15
-
09:00
-
08:45
-
08:30
-
08:15
