Claude Just Started Signing Everything It Writes

An invisible mark, no opt-out, and a catch, it proves Claude touched your text, not that Claude wrote it.

Welcome Back,

Anthropic will embed invisible watermarks in text produced by its new Claude models and include signed provenance metadata with generated files, according to reporting from Search Engine Journal and multiple outlets covering an updated Anthropic support article. The marking is implemented worldwide, not just in the European Union, and applies with no opt-out.

The trigger is Article 50 of the EU AI Act, which requires generative AI providers to mark outputs in machine-readable formats so downstream users, regulators, and platforms can determine what content is AI-generated. Those transparency obligations became applicable on August 2, 2026, and Anthropic signed the Act's Code of Practice on Transparency of AI-Generated Content as a provider of both generative AI models and generative AI systems, joining Google, Meta, Microsoft, Mistral, and OpenAI on that same provider list.

Today we look at exactly how the watermark works, what it actually proves and doesn't prove, why it can survive copy-paste but not heavy rewriting, and what's fueling real objections from writers and researchers who feel caught by a system they never agreed to.

🔏 How the Mark Actually Gets Made

Text gets an embedded watermark that Anthropic says doesn't change the meaning, quality, or readability of a response. According to Forbes' coverage, the text watermark subtly biases Claude's word choices, making patterns detectable over enough content, and travels with copied text, so pasting a Claude response into another document doesn't strip the signal.

Files work differently. Generated files can receive signed provenance metadata based on the Coalition for Content Provenance and Authenticity, or C2PA, standard, an open standard covering supported formats including .svg, .png, and .jpg. A valid C2PA manifest signals that Claude processed the file and lets a viewer detect whether that metadata has since been altered.

That model-level implementation means the marks apply across every surface people actually use Claude through, the API, Claude Code, Claude Cowork, Claude Tag, and deployments through AWS, Google Cloud, and Microsoft Foundry. Claude models launched on or after August 2, 2026 support marking at launch. Older models are being retrofitted, with no firm timeline announced yet.

The watermark is woven into word-choice patterns across a response, not added as a visible tag or footer.

❓ "Processed" Is Not the Same as "Authored"

Here's the single most important caveat in this whole story, and Anthropic states it directly. According to Anthropic, a detected mark indicates that Claude may have processed the content, but it does not confirm that Claude authored it.

The reverse is just as true, and just as important. The absence of a mark is not definitive confirmation of human authorship either. There will be no detectable signal on content from older, pre-marking models, very short passages, heavily paraphrased text, or files whose metadata was stripped through screenshots or format changes. A completely AI-generated essay could easily carry no detectable mark at all.

🧩 What Actually Breaks the Watermark

Because the mark is a statistical signature spread across word-choice patterns, it's genuinely resilient to some things and genuinely fragile against others, and understanding that split matters more than any single headline claim about "removal."

Copying and pasting text into a new document doesn't remove the mark, and neither does light editing, format changes, or moving text between platforms. What does affect detectability is heavy editing, paraphrasing, translation, or mixing Claude's output substantially with other writing, along with very short passages that simply don't contain enough text for a reliable signal. C2PA file metadata is even more fragile, any screenshot, format conversion, or social media re-upload strips it, and open-source removal tools for that layer already exist.

The EU AI Act's transparency rules apply to any provider offering generative AI in the EU, but Anthropic chose to apply its marking globally.

📜 The Law Actually Behind This

Article 50 of the AI Act requires that outputs of generative AI systems be "marked in a machine-readable format and detectable as artificially generated or manipulated," using technical solutions that are "effective, interoperable, robust and reliable as far as this is technically feasible." The Code of Practice regulators expect providers to follow calls for a multi-layered marking strategy, since no single technique is considered sufficient on its own, which is exactly why Anthropic paired an invisible text watermark with separate C2PA file metadata rather than relying on just one method.

That's a genuinely important legal nuance worth understanding. Anthropic's job under the law is to make Claude's output technically detectable as AI-generated. Whether an end reader ever sees a visible "AI-generated" label on a news article or a piece of marketing copy is a separate obligation that falls on whoever actually publishes that content, not on Anthropic.

😠 Why Some Users Are Genuinely Upset

This policy has drawn real, substantive objections, not just background grumbling. Several lawyers, academics, researchers, and writers who use Claude to copy, edit, or revise work that's substantially their own now face the prospect of that work carrying an AI marker they never agreed to and can't opt out of.

Anthropic isn't the first to weigh this trade-off, and not everyone landed in the same place. OpenAI scrapped its own text watermarking plans back in September 2024, after a company survey found almost 30% of ChatGPT users said they'd use the product less if watermarking were added, a genuinely direct illustration of the tension between transparency obligations and user trust that every major lab is now navigating differently.

Writers who use Claude only to lightly edit their own work now have to weigh a marker they didn't ask for.

🧠 AI Spotlight Analysis

The most useful way to understand this system is as a weak positive signal with essentially no negative signal at all. A detected mark tells you something, Claude touched this content, but it says nothing about who authored the ideas, how much of the text is actually machine-written, or whether the result is any good. It's built to make provenance checkable at scale across billions of interactions, not to adjudicate any single dispute over whether a specific document was written by a person or a model.

That's precisely why treating a watermark hit, or a miss, as a verdict in something like an academic integrity case or a publishing dispute would be a genuine misuse of the tool. Anthropic's own framing, "may have been processed," not "was authored," is a careful, honest hedge, but hedges only protect people who actually read them before drawing conclusions.

Anthropic is not alone in this space, Google already uses SynthID to embed invisible watermarks in AI-generated text, and OpenAI currently uses transparency tools including SynthID for images and audio but hasn't publicly announced a text detection system of its own. What sets Anthropic's move apart is watermarking the format Claude is most used to produce, and the hardest to mark reliably, plain text, and framing the work explicitly as legal compliance rather than a voluntary provenance gesture.

🔗 Sources and Further Reading

Thanks for reading AI Spotlight.

Our mission is simple: deliver clear, trustworthy, and actionable AI insights that help professionals stay ahead without the hype.