Claude AI Watermarks Text. China Made It Law First.

Anthropic announced on August 11 that Claude models launched on or after August 2 now weave an invisible watermark into every piece of text they generate. It's spreading on social media as news, and a lot of the reaction treats it as Anthropic inventing something. The actual chronology runs the other direction. China turned this into enforceable law in September 2025, with real penalties attached, nearly a year before Claude did it voluntarily.

What Claude Actually Does

The mechanism is a statistical watermark, not a visible tag. When a supported Claude model generates text, it biases its word choices in a pattern that's invisible to a human reader but detectable by software built to look for it. It travels with the text through copy and paste, and Anthropic says it can survive some editing, though heavy paraphrasing, translation, or reprocessing by another AI model tends to break it. It applies at the model level, worldwide, across the API, Claude, Claude Code, Cowork, and Claude Tag, with no opt-out. Anthropic is also attaching signed C2PA provenance metadata to generated files like PNGs, JPGs, and SVGs.

There's a real, disclosed cost to this. Alexander Cui, head of research at GPTZero, told Tom's Guide the watermarking forces the model into statistically less common word choices to embed the pattern, and confirmed directly: "The output quality does subtly decrease." Anthropic is also careful about what a detected mark actually proves, and it's less than people assume. A watermark means Claude processed the text, not that Claude wrote it. Someone who used Claude only to proofread, translate, or reformat their own writing gets the same mark as someone who had Claude generate the whole thing from scratch. And a missing mark proves nothing either, short passages, heavy edits, and older pre-August models won't carry a detectable signal at all. An Anthropic engineer put the honest limitation plainly in a follow-up: "it's not perfect, you can edit it, but it's a first step."

That imprecision is exactly what's driving the backlash on X and Reddit, mostly from people who use Claude for light editing on their own writing and don't want it flagged as AI-generated, and from developers uneasy about a signal embedded in code they can't fully control. The driver behind the timing is specific too: the EU AI Act's Article 50 transparency requirement, which is why Anthropic is also retrofitting marking support onto some pre-August models during the law's transition period.

China Made This Law, Not a Policy, Almost a Year Earlier

Here's the part missing from most of the social coverage. China's Cyberspace Administration released the Measures for Labeling AI-Generated Content on March 14, 2025, paired with a mandatory national technical standard, GB 45438-2025. It took legal effect on September 1, 2025. That's not a company policy a lab can quietly revise. Non-compliance carries regulatory investigation, fines, business suspension, and license revocation for repeat offenders. And it didn't appear from nothing, it builds directly on 2023 rules governing "deep synthesis" content, meaning the regulatory groundwork for mandatory AI watermarking in China predates Claude's version by roughly three years, not eleven months.

The Chinese requirement is also structurally broader than what Anthropic just shipped. It mandates both an explicit label, a visible notice telling the user the content is AI-generated, and an implicit label, a hidden watermark or metadata tag, across text, images, audio, video, and synthetic scenes. Claude's current implementation is entirely the hidden half, there's no visible in-chat notice telling a user their output was watermarked.

The compliance is documented, not theoretical. DeepSeek appends a visible disclaimer, "content generated by AI, for reference only," alongside hidden technical markers embedded in metadata, and explicitly prohibits users from altering or removing either. Zhipu AI's Qingyan chatbot displays an "AI-generated" notice directly on responses, with a disclaimer that follows the text when it's copied, and the company applies both explicit watermarks and hidden digital identifiers across text, images, and video. SenseTime tags its digital humans and appends disclaimers to chatbot conversations, and has published APIs so partners can integrate the same labeling. The law applies to any AI service provider operating in China, so it isn't optional for Alibaba's Qwen either, though a specific public compliance statement from Qwen by name isn't something this piece can point to directly.

The Rest of the Timeline the Headlines Are Skipping

Two more data points make the "Anthropic did something new" framing even harder to sustain. Google has been watermarking AI-generated images since 2023, through SynthID, and has since expanded the same approach to text, audio, and video. That's the longest continuous run of any major lab, Western or Chinese, predating even China's formal law.

OpenAI is the genuinely interesting holdout, and not because it lacks the technology. According to the Wall Street Journal, OpenAI has had working text-watermarking capability ready for ChatGPT for years and has deliberately chosen not to release it. The internal debate has run for years, with concerns about false positives, how easily the signal could be circumvented, and fear that shipping it would push users toward competitors without the same friction. OpenAI built the tool and sat on it. Anthropic built a similar tool and shipped it.

Lay the real chronology out in order: China's regulatory groundwork starts in 2023 and becomes enforceable law in September 2025. Google's been marking images since 2023. OpenAI has had text watermarking ready for years and hasn't shipped it. Claude's version arrived in August 2026, the most recent of the group, and the first major Western lab to do it without a government mandate forcing the date. That's a real, specific, and fairly flattering position for Anthropic to be in, voluntary rather than legally compelled, just not the "Claude invented this" story circulating right now. It's closer to being last to a table China set nearly a year ago.

Previously on TheQuery:

• The US Government Is Now Approving AI Models Before They Ship
• Kimi K3 Is the Largest Open-Weight Model Ever Announced