Anthropic announced last week that all the text Claude generates now carries an invisible watermark.
This watermark hides in the statistical patterns of the word choices themselves, rather than tacking a "this text was generated by AI" line onto the end of every response. It reads exactly like normal text to us, but a dedicated detection tool can pick it out by recognizing those patterns.
So what does a text watermark like this actually mean? Copy a paragraph Claude wrote into an email, paste it into a Word document or a school's essay system, and the watermark is still there. That is different from the image watermarks you are used to, which vanish the moment you take a screenshot or crop the picture. Text watermarks do not work that way; they live inside the content itself.
A bit more on how image and video watermarks differ. Take Google's SynthID (https://deepmind.google/models/synthid/), the most widely used approach right now, which had already watermarked over 10 billion pieces of content by May this year. It trains two deep learning models together: one embeds an invisible watermark directly into the pixel values of an image, and the other detects it.
Video gets watermarked frame by frame, so trimming the clip alone will not remove it. People have tried adding filters, changing the frame rate, and compressing the quality, and none of it strips the mark. Run it through a detection tool and it returns one of three verdicts, "watermarked," "uncertain," or "not watermarked," plus a confidence score. SynthID is already built into all of Google's products.
There is another approach called C2PA (The Coalition for Content Provenance and Authenticity). Instead of weaving the watermark into the content, it attaches a cryptographically signed record of the content's origin to the file's metadata. The upside is that it is an open, unified standard and traceable; the downside is that a screenshot or a format change wipes that information out. So the pixel-embedded SynthID is far more robust than anything hung on metadata.
Where are the verification tools at right now? We touched on this last time when we talked about GPT-Live.
- OpenAI: has a public Verify tool (https://openai.com/research/verify/) where you upload an image or an audio clip and it tells you whether there is a C2PA or SynthID watermark inside. In the July 31 update, besides adding audio, they also started opening up a verification API so developers can build this provenance check into their own products.
- Google: offers the SynthID Detector, built specifically to detect SynthID watermarks. But I have not seen a GUI entry point like OpenAI's. API is here:
- Anthropic (Claude): the text watermark is announced, but I have not seen an API for it yet.
But this text watermark is not bulletproof either. The technology has clear limits.
Heavy rewriting, running it through a translator into another language, or repeated manual polishing can all dilute the signal to the point where it cannot be detected. Very short passages do not carry a strong enough mark and may not be detectable at all.
That said, the watermark only proves Claude "touched" the text, not that Claude wrote it from scratch. Hand Claude your own draft to tidy up and the result gets marked all the same. So if a detection result says "AI-generated," it only tells you AI was involved, not to what degree.
So What? These days everyone is inventing all sorts of things in a hurry, and stacking equally over-the-top compliance checks on top. A document that does not get flagged is not necessarily written by a human; a document marked as AI is not necessarily fully ghost-written by one.
Anyway, I would honestly love to see how many books published after 2023 were written by AI, haha.