You can't see it. But it's there.

On August 17, Anthropic announced: future versions of Claude will include invisible text watermarking technology to meet EU AI Act compliance requirements.

The key detail is how this watermark works—it's not "inserting invisible hidden characters into text" as many imagine, nor extra tags or markings, and it does not contain user identity information. It's based on statistical patterns in the model's token selection—when the model picks the next token, it follows a statistical pattern known only to Anthropic. The resulting text looks completely normal, but Anthropic can detect via statistical analysis whether it was generated by Claude.

Why This Matters

First, it signals that EU AI Act enforcement is entering a substantive phase.

The EU AI Act is the world's first comprehensive AI regulation, with high-risk AI system provisions taking full effect in August 2026. One core requirement: AI-generated content must be identifiable—whether text, image, video, or audio, users have the right to know if what they're seeing is AI-generated.

Anthropic isn't the first company to do watermarking, but it's the first major LLM vendor to clearly explain the technical details and commit to not including user identity information. More importantly, Anthropic explicitly stated: other major model developers have signed the same Code of Practice and will implement watermarking in lockstep.

What does that mean? It means content generated by GPT, Claude, Gemini—all will carry their own "statistical fingerprints." You won't need "AI-generated" labels or guesswork from detection tools; statistical analysis can accurately trace provenance.

This is certainly positive for combating deepfakes, AI disinformation, and AI scams—fake news, fake videos, and fake voices can be traced back to the model vendor, clarifying liability. But it also raises privacy and free speech concerns: if all AI-generated content is traceable, will people feel inhibited when using AI to assist with writing, creativity, or expression? Will it lead to self-censorship?

The Era of AI Content "Birth Certificates"

The spread of invisible watermarking means AI content officially enters an era of "licensed to operate."

The old internet was "anonymous by default"—content you posted online was hard to trace back to a specific person or tool unless you actively signed it. Future AI content will be "traceable by default"—every AI-generated text, image, and video will carry the model vendor's watermark, like a factory nameplate on a product.

Is this change good or bad? There's probably no simple answer.

On the positive side, it can effectively curb AI misuse—deepfakes, AI disinformation, AI-written spam, AI-generated scam content will see some relief. Regular users will have a much lower barrier to identifying AI content and won't be fooled by AI-generated misinformation.

On the concern side, it may limit legitimate uses of AI tools—for example, an author who uses AI to brainstorm, draft, and polish but whose final work is original—is that considered AI-generated? Does it need a watermark? Will detection tools produce false positives? These questions will need more nuanced这些问题将需要更细致的规则来界定。

But whether you like it or not, the era of AI content provenance is here. Just as the internet moved from "anonymous" to "identified" as an inevitable trend, AI content moving from "indistinguishable" to "traceable" is also irreversible. Anthropic's announcement is just the first formal footprint in this trend.

See you tomorrow.