Anthropic Will Hide Secret Watermarks in All Claude Text and Images
The company behind the Claude AI assistant is embedding invisible markers into every response, so platforms and individuals can tell if content was machine-made.
Key points
- Anthropic has announced it will embed invisible watermarks in all text and images generated by Claude, its AI assistant.
- The move is driven by the EU AI Act, the European Union's new law on artificial intelligence transparency, which took effect on 2 August 2025.
- Existing Claude models will be updated gradually; new models will carry watermarks from launch day.
- Images will use C2PA, an industry-standard labelling system already used by Adobe, OpenAI and Google.
- Anthropic has not yet named its text-watermarking technology, and experts note both methods have known weaknesses.
If you have ever read an article, a social media post, or a customer email and wondered whether a human or an AI wrote it, Anthropic wants to give you a technical answer. The company behind Claude, one of the most widely used AI assistants, announced on a new support page that it will embed invisible, machine-readable markers into every piece of text and every image Claude produces.
Those markers are invisible to human eyes. Platforms, browser tools, and detection services will be able to read them automatically.
What exactly is being added, and how does it work?
Two separate techniques cover two types of content. For images, Anthropic will attach C2PA metadata, a kind of digital certificate of origin that records where a file was made. Think of it as a nutritional label sewn into the lining of a garment: you cannot see it, but a scanner can. Adobe, OpenAI, and Google already use C2PA on their AI image tools.
For text, Anthropic is using what it calls an "imperceptible watermark" woven directly into the words Claude writes, without changing meaning or readability. Anthropic has not named this system or published technical details yet, saying it will release documentation soon. Crucially, the company says the watermark travels with the text when someone copies and pastes it, and may survive light editing.
The markers will apply across every Claude surface: the Claude website, Claude Code (its coding assistant), and access through cloud services such as Amazon Web Services, Google Cloud, and Microsoft Azure.
Does this mean AI-written content will always be detectable?
Not reliably, no. Anthropic itself acknowledges the limits plainly: content that lacks detectable marks could still have come from a generative AI tool. Both techniques have known gaps.
C2PA metadata is easily stripped, sometimes accidentally, when a file is uploaded to a social platform that re-processes images. Text watermarks have not been independently tested here because Anthropic has not yet published how its system works. Even the best watermarking research to date shows that heavy editing or paraphrasing can degrade or erase a watermark entirely.
| Content type | Marking method | Who else uses it |
|---|---|---|
| Images | C2PA provenance metadata | Adobe, OpenAI, Google |
| Text | Unnamed imperceptible watermark | Not disclosed |
| Existing Claude models | Gradual rollout (timeline TBC) | N/A |
| New Claude models | Watermarked from day one | N/A |
First reported by The Verge AI, the announcement positions Anthropic ahead of an EU deadline rather than behind it. The EU AI Act includes a four-month grace period for AI products that already existed before 2 August 2025, so Anthropic has time to update its current models. Detection tools for reading these watermarks do not yet exist publicly; Anthropic says they are coming.
What does this mean for ordinary people?
For now, not much changes in daily life. The markers are invisible and there is no consumer tool to check them yet. Once detection is available, readers, journalists, teachers, and platform moderators will have a more reliable way to flag AI-generated content than today's guesswork.
If you receive a document, a news post, or a business email you suspect was AI-written, the advice for now is the same as before: read critically, check sources, and wait for the detection tools Anthropic has promised.
Common questions
Will this stop AI-written content from spreading online?
No. Watermarks make detection easier for those looking for them, but they do not prevent Claude from generating content or stop anyone from sharing it.
Can someone remove the watermark?
Possibly. C2PA image data is known to be stripped by many upload processes. Text watermarks may degrade under heavy editing. Anthropic has not published enough detail yet to say how durable its text system is.
