Anthropic is adding invisible watermarks to Claude’s AI-generated text that can be detected even after you copy and paste it

Is this the beginning of the end of AI-generated text as we know it?

Since the launch of ChatGPT nearly four years ago, the volume of AI-generated text has exploded, fueling predictions that machine-generated content could eventually surpass human-created text across the internet. One of the defining features of that growth has been how easily AI writing can leave the chatbot and blend into everything else online. Copy a response from Claude, paste it into an email, article, document, or website, and the connection to the AI that generated it could effectively disappear.

Anthropic is about to challenge that assumption.

Copy text from Claude and paste it somewhere else, and something invisible could now come along with it.

Anthropic announced today that it is adding machine-readable watermarks directly to text generated by supported Claude models, creating an invisible signal that can travel with the words after they leave Claude. The company says the watermark does not alter the meaning, quality, or readability of the response and may survive some editing.

The move is part of Anthropic’s implementation of transparency requirements tied to the European Union’s AI Act. Yet its reach goes far beyond Europe. Anthropic says the markings will apply to supported Claude models wherever they are offered worldwide.

“New models will mark AI-generated content from day one,” Anthropic said. “Claude models launched in the EU on or after August 2, 2026 will support machine-readable marking at launch. Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported.”

The markings will extend across Claude Platform, Anthropic’s API, Claude, Claude Code, Claude Cowork and Claude Tag. Supported Claude models accessed through AWS, Google Cloud or Microsoft Foundry will carry embedded text watermarks too, according to Anthropic.

That makes this more than a label tucked inside the Claude interface. Anthropic says the watermark is applied at the model level, allowing it to remain part of generated text regardless of which supported Claude product produced it.

An invisible watermark that follows the text

Anthropic is using two different methods to mark content. Text receives an embedded watermark. Supported files, including PNG, JPG, and SVG files, can receive digitally signed provenance metadata based on the Coalition for Content Provenance and Authenticity, or C2PA, open standard.

“When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself,” Anthropic said. “You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response.”

The copy-and-paste behavior is what makes the system particularly interesting.

Anthropic says new Claude models will embed machine-readable watermarks in AI-generated text and that it will provide ways for users and third parties to detect Claude’s marks. Credit: Anthropic

“Because the watermark is part of the text, it will travel with the text when it’s copied and pasted elsewhere, and may persist through some editing,” the company said.

Anthropic plans to give users and third parties a way to detect its marks. Technical details about those detection mechanisms are still forthcoming, leaving one major question unanswered: exactly how Anthropic is embedding the watermark into generated language.

There are limits, and they matter.

Finding a Claude watermark does not prove Claude originally wrote the material. A person could write something themselves and use Claude to proofread, translate, summarize, or reformat it. The resulting text could contain Claude’s mark despite the ideas and original writing coming from a human.

The opposite is true too. Failing to detect a watermark does not prove something was written by a person.

Anthropic says heavy editing, paraphrasing, translation, mixing Claude output with other writing, or using passages that are too short could make the watermark difficult or impossible to detect. Older Claude models may lack marking support. File provenance can disappear through screenshots, format conversions, or other processes that strip metadata.

Those caveats distinguish Anthropic’s approach from the controversial AI-writing detectors that attempt to guess whether text came from a machine. Anthropic is creating a signal inside its own output rather than asking a detector to infer authorship solely from writing patterns.

The policy stems from Anthropic signing the EU AI Act’s Article 50(2) Code of Practice on Transparency of AI-Generated Content. Claude models launched in the EU on or after August 2, 2026 are covered from launch, and Anthropic says it is working to bring marking support to models released before that date.

For millions of people using AI-generated text, the change could make one of computing’s most ordinary actions far more consequential. Copying Claude’s words may soon mean copying something else too: an invisible trail showing that the text passed through AI.

The fallout: How Anthropic’s invisible watermarks could upend the AI detection industry

The bigger fallout may be felt outside Anthropic. An entire industry has emerged around AI-text detection, with companies selling software to schools, publishers, employers and other organizations that want to determine whether writing was produced by AI. Those systems typically analyze text and estimate the likelihood that a machine generated it.

Claude’s watermarking approach could change that equation. Instead of asking a third-party detector to infer whether Claude was involved, users could eventually check for a machine-readable signal deliberately embedded by Anthropic itself.

That distinction becomes especially significant for organizations whose standard is that submitted work must be entirely human-written without generative AI assistance. Anthropic cautions that detecting its watermark does not prove Claude originally authored the material. Claude could have proofread, translated, summarized, or edited human-written text. Yet under a strict no-AI policy, that distinction may matter less. If Claude processed the final text, AI was involved.

The immediate threat to AI-detection companies is limited by the watermark’s own weaknesses. Anthropic says heavy editing, paraphrasing, translation and other changes can make the signal undetectable, and older Claude models may lack marking support. Detection companies could still analyze unmarked text or evolve into services that check multiple provenance systems.

The bigger question is what happens if Anthropic is just the beginning. If OpenAI, Google, Meta and other major AI providers adopt comparable machine-readable text markings, the business of detecting AI writing could shift from probabilistic guesswork to checking signals placed there by the model makers themselves.

That could force today’s AI-detection companies to reinvent themselves as provenance platforms rather than AI-writing detectors.

For nearly four years, the internet has treated copied AI text much like any other text. Anthropic is betting that the words can carry a history with them. If that idea becomes an industry standard, the consequences could reach far beyond Claude.