Claude is leaving its invisible mark, and the internet has questions

Anthropic is preparing to watermark text generated by future Claude models, introducing an invisible method of identifying the likelihood that its AI was involved in producing a piece of content as new European Union transparency requirements come into force.

The AI company said the watermark will be applied globally at launch and will not add hidden characters, identifying information or additional tokens to Claude's responses. Instead, it will create a detectable pattern through the word choices the model makes while generating text, allowing those with the appropriate key to assess the likelihood of Claude's involvement.

The move comes as Anthropic looks to comply with the EU AI Act. The company was among roughly 190 signatories to the EU Code of Practice on Transparency of AI-Generated Content in July 2026, which requires AI system providers to adopt methods for marking AI-generated text.

Don't miss: When Claude makes anyone a 'designer', what happens to creative craft?

The disclosure has already sparked significant public interest, with much of the conversation focusing on whether invisible watermarking can hold up once AI-generated text is edited, copied or reused. Since the revelation, media intelligence company Truescope tracked more than 7.1 million views around the conversation across Tier-1 TikTok, YouTube and Instagram, alongside 364,000 likes, 21,300 comments and 27,600 shares.

The most engaged discussions centred on watermark removal tools, authorship and attribution of AI-assisted work, the EU AI Act's global reach and the risk of human work being incorrectly labelled as AI-generated.

Some users backed watermarking as a necessary accountability measure and a step towards greater transparency, while others questioned whether it would be technically resilient. Circumvention became a particularly prominent thread, with an open-source GitHub repository called "watermarks-remover" by developer Guillaume Meyer surpassing 13,600 stars as of 18 August, according to Truescope.

CARMA's analysis similarly found that reaction leaned more cautious than enthusiastic. While 69.7% of sentiment was neutral, negative sentiment stood at 25.1%, compared with 5.1% positive, suggesting the announcement generated more questions and analysis than outright approval.

Much of that scrutiny focused on how the technology would work in practice. "Invisible" appeared in 45.3% of conversations analysed by CARMA, while "editing" appeared in 26.1%, "detection" in 19.6% and "provenance" in 18.7%. "Code" also featured in 34.6% of conversations, pointing to interest in how watermarking would work beyond conventional written content.

The conversation also continued to build after the initial disclosure. CARMA found that discussion rose from between zero and two posts per interval before the announcement to multiple spikes after 11 August, peaking at 70 posts on 17 August.

How the watermark works

At the heart of the debate is how Claude's invisible watermark actually works. The system will use a version of SynthID-Text, a technique developed by Google DeepMind and detailed in a Nature paper in 2024. Rather than inserting something into a response after it has been generated, the approach alters the source of randomness Claude uses when choosing between several similarly suitable words.

For instance, when several words could naturally complete a sentence without changing its meaning or quality, the watermark helps determine which option Claude selects. Across a sufficiently long passage, those decisions form a pattern that can later be checked against a key.

According to Anthropic, readers will not be able to distinguish watermarked text from text produced without the system. Its internal testing found no impact on the content, creativity or readability of Claude's responses, while the process is expected to have a negligible effect on model speed and no impact on the price of using Claude.

The watermark will also carry no information identifying an individual user, organisation or conversation. Anthropic stressed that detection would only indicate the likelihood that Claude was involved in producing or processing a piece of text, rather than proving that Claude was solely responsible for writing it.

There are limits to the technology. Detection becomes more reliable as passages get longer, as more word choices give the watermark room to form. Factual passages, proofreading and code may carry weaker signals because there are fewer instances where Claude can freely choose between equally valid alternatives without affecting accuracy.

This also means text that Claude has only lightly proofread may not contain enough of the watermark to reliably indicate its involvement. Translations, however, will be watermarked because Claude is responsible for choosing the words throughout the output.

The system is also not tamper-proof. Anthropic acknowledged that while light editing is unlikely to remove a watermark entirely, substantially rewriting a passage could do so. The watermark also cannot determine whether content was human-written or produced by another AI system.

Anthropic plans to introduce a watermark detection API, although it has yet to disclose details on how the tool will be implemented.

For images and other supported files such as PNG, JPG and SVG files, Claude will take a different approach. Rather than using the text watermarking system, the company will attach C2PA content credentials to file metadata, providing a cryptographically signed record that the file was created or processed using Claude.

Anthropic said it is initially rolling the watermark out globally because it does not yet have a durable way to restrict the technology by region. Older Claude models launched before 2 August 2026 fall under a transition period in the EU rules, with Anthropic working to introduce watermarking to those models over the coming months.

The watermarking push also comes as Anthropic expands Claude's role in creative and content production. Earlier this year, the company unveiled Claude Design, an AI-powered product aimed at turning visual creation into a conversational process, from early concepts through to production-ready assets.

Built on Claude Opus 4.7, the tool allows users to generate prototypes, presentations, marketing assets and landing pages by describing what they need. Outputs can then be refined through prompts, inline edits and adjustable controls, bringing multiple stages of the design workflow into a single interface.

Related articles: 
Anthropic builds Singapore footprint with hiring push   
Anthropic files confidentially for proposed IPO   
Anthropic appoints Theo Hourmouzis to lead ANZ as Sydney office officially opens