If you’ve used Claude, ChatGPT, or any other AI assistant recently, there’s a decent chance some of what it wrote you now carries an invisible signature, one you can’t see, but that a computer can check for later.
Anthropic, the company behind Claude, announced in August that it’s rolling out a watermarking system across its AI models. The idea is simple, even if the execution isn’t: every time Claude writes something, it has to choose between several equally reasonable words at different points in the sentence. Normally that choice is random. Now, it’s guided by a secret key, which creates a subtle, consistent pattern running through the text. Nobody reading it would ever notice. But if you had the key, you could check whether a piece of writing was likely to have come from Claude.
It’s not visible text, hidden characters, or anything you’d stumble across by accident. It lives in the statistical texture of the words themselves.
Why now? The honest answer is regulation as much as good intentions. Anthropic has said plainly that it’s implementing this to comply with the EU AI Act, and it signed onto the EU’s voluntary Code of Practice on transparency for AI-generated content earlier this year, alongside roughly 190 other companies. It’s part of a broader shift where AI-generated content is expected to be identifiable, not just convincingly human.
Here’s what it actually means for you, in practical terms. This isn’t a lie detector. Anthropic has been upfront that the watermark can only estimate the likelihood that Claude was involved. It can’t prove a piece of writing is entirely AI-generated, and it can’t tell you anything about other tools like ChatGPT or Gemini. It also has real blind spots: short posts don’t carry enough word choices to detect reliably, factual writing (think dates, formulas, code) leaves little room for the system to work with, and a thorough human rewrite can erase the pattern almost completely. Light editing probably won’t remove it — but a genuine rewrite will.
So if you were hoping watermarking would let you definitively sniff out AI-written comments, reviews, or forum posts going forward, temper that expectation. It’s one signal among many, not a verdict.
What it does do is add a small, quiet layer of accountability to a space that’s had very little of it. And it’s a reminder of something worth sitting with regardless of the technology: genuine, lived experience, the kind that can’t be reverse-engineered from a statistical pattern, is still the thing people are actually looking for when they come online to feel less alone. No watermark, present or absent, changes that.