You’ve probably seen the announcement by now. Anthropic says Claude now weaves an “imperceptible watermark” directly into its generated text. You won’t see it. It doesn’t change meaning, quality, or readability. It’s just… there. Invisible. Unverifiable. And entirely dependent on you believing Anthropic’s word.
Let’s sit with that for a second.
An invisible watermark is indistinguishable from no watermark at all.
That’s not a gotcha. That’s not me being contrarian for clicks. That’s basic epistemology. If a feature is, by design, undetectable to the human eye and — based on what Anthropic has shared so far — undetectable by any independent third-party tool, then the watermark itself is not the signal. The announcement is the signal. The press release is the product.
And that should make you deeply uncomfortable.
Here’s the problem: we’re entering an era where distinguishing between human-written and AI-generated text is becoming existentially important. Journalists need to know if a source document was fabricated. Teachers need to know if a student wrote their essay. Employers need to know if a candidate’s portfolio is real. The entire credibility infrastructure of written communication is at stake.
Anthropic steps in with a solution: we’ll mark our text. Great. Except the mark is invisible. And there’s no public tool to detect it. And no independent verification process. And no explanation of how it actually works.
John Gruber nailed this on Daring Fireball. He quoted Anthropic’s own language — “it weaves an imperceptible watermark directly into the text itself” — and pointed out the obvious: either there’s some non-printable Unicode trickery going on, or this is a promise wrapped in vapor. Either way, the user is asked to accept something they cannot test.
They’re not selling detection. They’re selling faith.
Think about what a watermark is supposed to do. A watermark on a dollar bill lets anyone hold it up to the light and verify authenticity. A watermark on a photograph lets anyone inspect the metadata. The entire point of a watermark is that it’s observable. It’s a trust mechanism that doesn’t require trust — you can see it with your own eyes.
Anthropic has inverted this completely. Their watermark requires MORE trust, not less. You must trust that the watermark exists. You must trust that Anthropic is applying it consistently. You must trust that it hasn’t been stripped or spoofed. You must trust that Anthropic’s detection tools (which you don’t have access to) are accurate. Every link in the chain is a leap of faith.
Now, let’s be fair. There are legitimate technical reasons why a watermark might need to be subtle. If it’s too obvious, bad actors will strip it. If it’s too heavy, it degrades output quality. The engineering challenge is real. I’m not dismissing the difficulty.
But here’s what I’m not going to do: pretend that an invisible, unverifiable, proprietary watermark is a meaningful step toward AI transparency. It’s not. It’s a corporate assurance dressed up as a technical feature.
The watermark isn’t in the text. It’s in the press release.
And that’s the real story here. Anthropic is doing what every AI lab does when cornered by the provenance problem: they’re performing responsibility. They’re creating the appearance of a solution without delivering the substance of one. The announcement generates headlines. The headline generates goodwill. The goodwill generates enterprise contracts. And somewhere in a server, a watermark that may or may not be detectable sits in text that may or may not be watermarked, verified by tools that may or may not exist.
This matters for you specifically if you consume, produce, or evaluate written content — which is to say, it matters for everyone reading this. Because every time an AI lab announces a transparency feature that you can’t independently verify, the bar for actual accountability moves a little lower. We get used to being told to trust. We stop demanding proof. And one day we wake up unable to distinguish a human voice from a machine’s, with nothing but a corporate blog post assuring us it’s all fine.
I’m not saying Anthropic is lying. I’m saying the structure of this announcement makes lying and truth-telling indistinguishable. And in a world where AI-generated text is flooding every channel of communication, that’s not good enough.
If Anthropic wants credit for transparency, they need to open the hood. Publish the detection method. Release a public verification tool. Let third parties audit the watermark. Let researchers try to break it. Let the community see whether it actually survives paraphrasing, translation, and adversarial stripping.
Until then, it’s a press release. Not a watermark.
FAQ
Q: If Anthropic says the watermark exists, why not just take their word for it?
A: Because the entire purpose of a watermark is to NOT require trust. A watermark on currency works precisely because anyone can verify it independently. If you need to trust the company's claim, it's not a watermark — it's a press release. The whole point is observable verification, and Anthropic has provided none.
Q: What does this mean for people who need to detect AI text in practice?
A: It means nothing changes. You still can't reliably tell if a piece of text came from Claude. The watermark is invisible, there's no public detector, and even if one existed, it would only catch text that wasn't stripped or paraphrased. For practical detection purposes, this announcement is a null operation.
Q: Isn't it unfair to criticize Anthropic when they're at least trying to address the provenance problem?
A: No. Performing responsibility without delivering verifiable results is worse than doing nothing, because it lowers the bar for what counts as 'addressing the problem.' It lets other labs point to Anthropic and say 'see, the industry is handling it' while nothing actually gets solved. Half-measures that look like solutions kill momentum for real ones.