Artificial intelligence-generated content could soon become easier to identify as Anthropic expands its invisible watermarking technology across more versions of its Claude AI models.
The company is extending the feature to older models as part of its efforts to comply with transparency requirements under the European Union’s AI regulations. The technology embeds an imperceptible digital pattern into AI-generated text, allowing specialised detection systems to assess whether Claude was likely involved in producing particular content.
The development comes as governments, educational institutions, publishers and businesses increasingly seek ways to distinguish between human-written and AI-generated material.
How Claude’s Invisible Watermarking Works
Unlike traditional digital watermarks that appear visibly on images or documents, Anthropic’s system operates through subtle patterns in the model’s word-selection process.
These patterns are embedded directly into the generated text and remain invisible to ordinary readers. According to the information provided, the watermark does not change the meaning, readability or overall quality of the content.
The technology is designed to make AI-generated writing identifiable without adding visible labels, symbols or unusual formatting to the final text.
One of its notable features is that the watermark can remain attached to content when users copy and paste it into other documents or online platforms.
It may also survive certain types of editing, meaning that making small changes to a paragraph may not necessarily remove the embedded signal.
Anthropic Expands Feature to Older Claude Models
Anthropic is extending the watermarking technology beyond its newer Claude systems to cover additional models.
The expansion includes Mythos 5, Opus 4.7, Opus 4.6 and Sonnet 4.6, with the wider rollout beginning on November 3, according to the report.
The move is intended to provide broader coverage across the Claude model family and make it easier for authorised organisations to identify content that may have been generated or substantially modified using the company’s AI tools.
By extending the feature to older systems, Anthropic is seeking to make its content-identification approach more consistent across different versions of Claude.
Watermarks Do Not Identify Individual Users
Anthropic says its watermarking technology does not contain personal information and cannot identify the individual user, organisation or conversation responsible for generating the text.
This distinction is important because the technology is intended to identify possible AI involvement rather than reveal the identity of the person who used the tool.
The system therefore focuses on the origin and processing of content, not on tracking individual users.
However, the presence of a watermark should not be treated as proof that an entire document was written by artificial intelligence.
Detection System Being Developed for Organisations
Anthropic is also developing a detection system that will allow eligible organisations to check whether Claude was likely involved in producing specific text.
Potential users include regulators, journalists, researchers and educational institutions.
Such a system could help organisations assess the origin of written material in situations where transparency is important, including academic submissions, professional documents and digital publishing.
For educational institutions, the technology could provide another tool for examining whether AI assistance was used in submitted work. Journalists and researchers may also find it useful when assessing the origins of digital content.
However, the detection process is expected to indicate possible involvement rather than deliver a definitive verdict about the authorship of an entire document.
Limitations of Invisible AI Watermarks
Despite its potential benefits, Anthropic’s technology has limitations.
The watermark may appear in text that was originally written by a human but substantially edited or translated using Claude. This means that its presence does not automatically establish that the original material was AI-generated.
Similarly, extensive rewriting or significant modifications can weaken or remove the watermark signal.
These limitations mean that organisations should avoid relying exclusively on watermark detection when making decisions about authorship or academic integrity.
Instead, the technology is best understood as an additional method of identifying possible AI involvement.
EU AI Act Increases Transparency Requirements
The expansion is linked to transparency requirements under the European Union’s AI Act, which seeks to make certain AI-generated content identifiable through appropriate marking methods.
As generative AI becomes more widely used, regulators are placing greater emphasis on transparency and accountability.
The rules are intended to help people recognise AI-generated or manipulated material and understand when artificial intelligence has played a role in creating content.
Anthropic’s approach represents one way of addressing these requirements through technology that operates in the background rather than displaying visible labels.
What This Means for Digital Publishing and Education
The expansion of invisible watermarking could have implications for several industries.
Digital publishers may gain another method of assessing whether AI tools were involved in preparing articles and other written material. Educational institutions could use detection systems as part of broader academic-integrity processes.
Businesses may also find the technology useful when reviewing documents and communications produced with AI assistance.
However, the effectiveness of these systems will depend on how reliably watermarks survive editing and how detection results are interpreted.
As AI-generated writing becomes increasingly common, Anthropic’s expansion highlights the growing effort to make AI involvement more transparent while preserving the convenience of generative AI tools.
The development also raises broader questions about authorship, content ownership and the distinction between human-written and AI-assisted material.
Also read: Anthropic Launches Claude Opus 4.7 as Its Most Intelligent Public AI Model
