Anthropic has released new details about how it plans to watermark text generated by Claude, explaining how the technology will identify AI-written content, how resistant it may be to editing and what impact it could have on code.
The company published a blog post outlining the system after its watermarking plans triggered debate among Claude users.
Anthropic says the change is being introduced as part of its effort to comply with transparency requirements under the European Union’s AI rules, which call for mechanisms that can help identify AI-generated content.
The company says the watermark will be invisible to ordinary readers and will not noticeably change the quality or style of Claude’s responses.
Watermarks Will Be Hidden in Word Choices
Anthropic said the system will work by influencing subtle choices Claude makes when several words or phrases would be equally appropriate.
For example, if a model could naturally choose between two similar words when describing something, the watermarking system could guide those choices in a particular pattern.
The resulting pattern would not be obvious to someone reading the text.
However, someone using an authorised detection system would be able to analyse the pattern and determine whether the content likely contains Claude’s watermark.
Anthropic stressed that the process is designed to leave the actual meaning and quality of the generated response unchanged.
SynthID-Text Technology Will Power the System
Anthropic said it plans to use the SynthID-Text approach developed by Google DeepMind.
Engineering & Technology
That technique was introduced as a way to embed detectable signals into AI-generated text without adding visible labels or changing the presentation of the content.
Anthropic also plans to provide a watermark-detection API, which could allow developers, platforms or other authorised users to check whether text carries the signal.
The company emphasised that this approach is different from conventional AI-detection software.
Traditional AI detectors typically analyse writing patterns, sentence structure or linguistic habits that appear common in machine-generated text.
Watermark detection instead looks for an intentionally embedded statistical pattern.
Users Debate the New Claude Feature
The announcement has generated strong reactions among some Claude users.
Online discussions have included criticism from people who worry that watermarking could reveal when they have used AI for work, education or other activities.
Others have defended the measure, arguing that users should be transparent when AI-generated material is presented as human-written work.
Reports have also suggested that some Claude subscribers have threatened to cancel their accounts because of the change.
Anthropic, however, has maintained that the watermark will not be visible in normal use and will not alter the quality of Claude’s output.
Editing May Weaken but Not Immediately Remove the Watermark
One of the biggest questions surrounding text watermarking is whether a user could simply edit the output and remove the signal.
Anthropic acknowledged that extensive rewriting could eliminate the watermark.
However, the company said minor revisions would probably not be enough to remove it completely.
If a user changes only a limited number of words or adjusts sentence structure while leaving most of Claude’s original response intact, enough of the watermark pattern may remain detectable.
A complete rewrite in which essentially every word is replaced could remove the signal.
Anthropic argued that once text has been rewritten that extensively, the question of whether it should still be classified as AI-generated becomes less straightforward.
Light Proofreading May Leave Little Detectable Signal
The system is also expected to behave differently when Claude is used primarily as an editor rather than as the original writer.
If a human writes most of a document and asks Claude to make only small grammatical or stylistic corrections, there may be very little AI-generated material available for the watermark to attach to.
Anthropic said detection in those circumstances will depend on how much of the original text Claude changes and how long the document is.
A heavily rewritten passage could contain a stronger watermark signal, while light proofreading might leave little or none.
That distinction could be important for people using Claude to improve their own writing rather than generate complete documents.
Code Will Be Less Affected
Anthropic also addressed concerns about programming output.
According to the company, computer code is likely to contain significantly less watermarking than ordinary prose.
That is because working code gives the model less flexibility in its choice of words, symbols and syntax.
Unlike natural-language writing, where several expressions may communicate the same idea, programming languages often require very specific structures to function correctly.
Anthropic said there may still be opportunities to apply watermarking in areas where wording choices are flexible, such as code comments or variable-related language.
However, it expects the effect on functional code itself to be negligible.
Watermarking Is Different From AI Writing Detection
Anthropic also sought to distinguish its watermarking system from tools that attempt to identify AI-generated writing based on stylistic characteristics.
Some AI-detection systems look for commonly repeated structures, phrases or writing habits associated with large language models.
Those systems make a prediction based on how the text appears.
A watermark, by contrast, is intentionally introduced during generation and detected using a corresponding technical key or system.
Anthropic argues that the two methods should therefore not be treated as equivalent.
Other AI Companies Expected to Introduce Similar Measures
Claude is unlikely to be the only major chatbot to adopt text watermarking.
Business & Corporate Law
Anthropic said other leading AI developers have committed to the same European transparency framework and are expected to introduce their own approaches for identifying machine-generated content.
That could make invisible watermarking increasingly common across the AI industry.
The broader goal is to create more transparency as AI-generated content becomes harder to distinguish from material written entirely by humans.
Transparency Debate Likely to Continue
Anthropic’s additional explanation may answer some technical questions, but the broader debate over AI watermarking is unlikely to disappear.
Supporters see it as a practical tool for improving transparency and reducing deceptive use of generative AI.
Critics worry about privacy, false assumptions and how watermark detection could affect legitimate uses of AI for editing, brainstorming and productivity.
For Claude users, the key takeaway is that the watermark is intended to be invisible, resistant to light editing and less pronounced in heavily constrained content such as software code.
As AI companies respond to new regulatory requirements, hidden markers of this kind could soon become a standard feature of generated text across multiple major platforms.
