Claude Now Watermarks What It Writes — But Not Your Code

TL;DR: Anthropic switched on watermarking for everything Claude writes, explained how it works on 14 August, and spent the weekend being shouted at about it. Here’s the bit for builders that nobody else is leading with: the Claude watermark barely touches working code, but it does mark the copy you write around it — your landing page, your product descriptions, your emails. Text you wrote yourself and only asked Claude to tidy stays essentially unmarked.

What actually happened

On 14 August 2026, Anthropic published an explanation of how the Claude watermark works. It isn’t visible, and it doesn’t add anything to what you get back. When Claude has a genuine choice between words that would work equally well, the watermark quietly biases which one it picks, using a secret key and the preceding few words. Feed enough of that text to a detector holding the key, and the pattern shows up.

Anthropic is doing this because it signed the EU Code of Practice on Transparency of AI-Generated Content in July, alongside roughly 190 other signatories. Notably, it’s applying the watermark worldwide, not just in Europe — its stated reason is that it doesn’t “yet have a durable way to scope it by region”.

Two days later John Gruber called it “a perversion of writing”, arguing that “no two synonyms carry the exact same meaning” and that “the idea that anything other than my needs should factor into the generation of text for me is patently offensive”. That post hit 188 points on Hacker News and set the tone for the weekend’s reaction. He wasn’t alone: TechCrunch reported Reddit threads debating whether the whole thing was a conspiracy, and — citing Business Insider — dozens of users on X saying they’d cancelled their subscriptions over it.

What the Claude watermark actually marks

This is the part worth your attention, and it’s more specific than the row suggests.

Watermarked	Largely exempt
Prose and creative writing	Code where an exact output is required
Translations (every word is Claude’s)	Constrained factual statements
Code comments	Text you wrote that Claude only proofread

Anthropic’s reasoning is mechanical rather than generous: the watermark needs choices to hide in. “Where an exact output is required,” it says, “the watermark isn’t applied” — working code has to be exactly right, so there’s little room to bias. Comments are prose, so they carry it.

The same logic covers the fear doing the rounds this weekend, that your own writing will get flagged. Anthropic’s position is explicit: when Claude proofreads something a person wrote, “nearly all the words are the person’s, [so] there’s very little (if anything) for the watermark to attach to.”

That’s Anthropic describing its own system, so treat it as a claim rather than an audited fact — but it’s a specific, testable claim, and it directly contradicts the loudest worry.

Three other details that got lost in the noise: detection doesn’t work well on short samples, only Anthropic can currently detect it, and a public detection API is coming with no date attached. There is no opt-out.

Why it matters if you don’t code

If you’re building an app with AI, you’re almost certainly writing with AI too — the landing page, the pricing blurb, the onboarding emails, the app store description. That’s the material the Claude watermark touches.

The practical shape of it:

• Your app’s code: essentially unaffected. Ship as normal.
• Your marketing copy, docs and emails: marked, if Claude wrote them.
• Your own writing, tidied by Claude: essentially unmarked, per Anthropic.
• Your code comments: marked, which matters to nobody but is true.

Whether that bothers you is a fair question with no universal answer. If you’re a solo founder writing your own product copy with an AI assistant, a detectable fingerprint is at worst a mild irritation. If you’re delivering client work on the understanding it’s human-written, this is a real consideration — and it’s a good moment to say that the honest fix there is telling the client, not defeating a watermark.

What to do

1. Nothing, for your code. The Claude watermark isn’t meaningfully in it, and no build or deploy step changes.
2. Assume AI-written copy is identifiable. Not by the public, and not today — but a detection API is coming. Write on that assumption rather than hoping otherwise.
3. If it matters, write the first draft yourself. Claude polishing your words leaves little to mark. That’s also, for what it’s worth, how you get better copy.
4. Tell clients if you use AI on deliverables. This was already the right call; it’s now the one that survives a detector.
5. Don’t go looking for a way round it. Anthropic hasn’t published one, “humanisers” are a low-quality industry, and the effort is better spent on the work.

Who should care (and who shouldn’t)

• Writing client-facing copy with Claude: the group this genuinely affects — decide your disclosure policy now.
• Building a revenue app: your code is fine. Only the words around it are marked.
• Building an internal tool: nothing to do at all.
• Learning, or not started yet: ignore this entirely. It changes nothing about which tool you should pick — the quiz is a much better use of the next 60 seconds.

Our take

The criticism is largely aimed at a different problem than the one builders have. Gruber’s objection is aesthetic and it’s a serious one — if you care about prose, a system nudging word choice for reasons unrelated to meaning is a genuine loss. We don’t think that’s the question for someone shipping an app.

Our reading is that the Claude watermark draws its line in a sensible place, if it holds: the stuff that must be exact is left alone, the stuff with room to spare carries the mark. The two things we’d keep an eye on are the detection API, which is where this stops being theoretical, and the fact that this is a global rollout driven by a European rule — a pattern worth noticing, because it won’t be the last time.

And a small note on where we started: Claude remains a strong choice for building things. This changes the copy you write, not the tool you should use.

Not sure which AI tool fits what you’re building? Take the 60-second Vibe Coding Tool Finder quiz →

Some links on this page are affiliate links. If you buy through them we may earn a commission at no extra cost to you.

FAQ