AI-generated content can no longer be hidden! Anthropic introduces a hidden watermark—here’s how it will be detected..
Anthropic is introducing a hidden watermarking system to identify text generated by its AI model, Claude. This means that if content is written using the Claude model without any subsequent editing, the system can determine that the content originated from Claude. Notably, this watermark will not introduce distinct characters, symbols, or visible alterations to the text. Instead, Claude will generate a specific pattern during word selection that the detection system can recognize. According to the company, the goal is to ascertain whether Claude played a role in creating the text.
**Hidden Watermark Created Through Word Choice**
According to Anthropic, when Claude selects one word from two or more equally suitable options, the system alters the source of randomness behind that choice. This creates a specific pattern throughout the text that a detection system can identify using a special key. Claude is not forced to use odd or unusual words in this process; Anthropic has utilized a version of Google DeepMind’s "SynthID-Text" technology (published in 2024) for this purpose. The company states that internal testing has shown no adverse effects on content quality, creativity, or readability.
**Not Every AI-Generated Text Will Be Fully Identified**
This watermark is designed to indicate the likelihood of Claude's involvement in a text, but it does not definitively prove the text's entire origin. It will not reveal the user's identity, their organization, or conversation details, nor will it indicate whether the text was generated by another AI. Detection reliability may be lower for short texts, as the system has less material to analyze regarding word choices. The watermark may also be less effective in factual writing, proofreading, and code generation, as these contexts offer limited scope for varying word choices.
**Complete rewriting removes it**
Anthropic states that while the watermark might persist after minor edits, it can be removed if the entire text is rewritten. The watermark will also remain in translations generated by Claude, as Claude itself selects the wording in these instances. The company distinguishes this method from traditional AI detection tools; while standard detectors analyze writing style and patterns, this technology identifies a watermark specific to a particular model. Anthropic is also developing a detection API that will allow users and third parties to verify the presence of this watermark in text.
**Use of C2PA credentials in image files**
In addition to text, Anthropic will utilize C2PA content credentials for supported files generated or processed by Claude. This includes formats such as PNG, JPG, and SVG. Cryptographically signed metadata will indicate Claude's involvement in the file's creation without revealing the user's personal information. C2PA is an open standard already employed by camera manufacturers and image editing software.
**Global rollout driven by the EU AI Act**
Anthropic is implementing this watermarking system globally, as the company currently lacks a reliable method to restrict it to specific regions. This move is also intended to meet the requirements of the European Union's AI Act. In July 2026, Anthropic signed the EU's Code of Practice regarding the transparency of AI-generated content—an initiative that included several major AI companies alongside approximately 190 other signatories. Anthropic is also working to add watermarks in the coming months to models launched prior to August 2, 2026.
Disclaimer: This content has been sourced and edited from TV9. While we have made modifications for clarity and presentation, the original content belongs to its respective authors and website. We do not claim ownership of the content.
