A free guide by New Money School
The headline moved fast this week, and most people read it as "AI can now catch you." That is not what was announced.
Here is the plain version: what the watermark actually is, what it genuinely can and cannot tell anyone, and the four limitations that almost every scary take left out.
What Actually Happened
Anthropic said that Claude models launched on or after August 2, 2026 will weave an imperceptible watermark directly into the text they generate. This is not a setting you turn on. It happens at the model level, which means it applies across the Claude apps, the API, Claude Code, Claude Cowork, Claude Tag, and Claude running on AWS, Google Cloud, and Microsoft Foundry.
It also applies worldwide, not just in Europe. Older Claude models are expected to get marking during the EU AI Act's transition period.
The line everyone is quoting is this one: because the watermark is part of the text, it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing.
The Part That Trips People Up
A watermark in an image is usually metadata, a tag attached to the file. Strip the metadata, the tag is gone.
This is different. The mark is inside the words themselves. When a model writes, it is constantly choosing between several words that would all work fine. A text watermark nudges those choices in a pattern that is invisible to you but detectable statistically across a long enough passage. Nothing about the meaning or the readability changes. There is no hidden character to find and delete.
That is why copy and paste doesn't remove it. You are copying the pattern along with the words.
Files work differently. For supported file types, Anthropic attaches signed provenance metadata using the C2PA standard, the same content-credentials approach camera makers and Adobe use. That part is metadata, and it behaves like metadata: re-saving a file through another tool, converting the format, or screenshotting it will usually strip it.
Four Things It Does Not Mean
This is the section worth actually internalizing, because almost every scary take skips it.
1. It is not proof you didn't write something. Anthropic said this outright. People use Claude to proofread, translate, summarize, and reformat their own work all the time. Run your own paragraph through Claude to fix the grammar, and the output can carry a mark, even though every idea in it is yours. A detected watermark says Claude touched this text. It does not say Claude wrote it.
2. No mark does not mean human. Text from a pre-August 2026 model carries nothing. Neither does output from any other company's model. Neither does a passage that got heavily rewritten. Absence of a watermark proves nothing at all.
3. It is statistical, not a fingerprint. Short passages don't carry enough signal to be reliable. A sentence or two is nowhere near enough. The confidence comes from length.
4. Heavy editing degrades it. Anthropic's own wording is "may persist through some editing." Substantial rewriting or paraphrasing weakens the signal. And for code specifically, running a formatter or linter will likely destroy token-level marks entirely.
Why They Did It
Anthropic signed the EU AI Act's Article 50(2) Code of Practice on Transparency of AI-Generated Content. That rule requires providers of generative AI to mark machine-made output so it can be identified as such. Rather than run one version of Claude for Europe and another for everyone else, Anthropic is applying it globally.
In other words: this is a compliance move, not a surveillance product. Nobody built a database of what you wrote. There is no lookup of your account. The mark is in the text, and detecting it requires having the text in front of you.
Detection
Anthropic has said it will publish technical details so people and platforms can check for its watermarks. At the time of writing, that tooling isn't broadly available yet, so treat any service claiming to definitively detect Claude text with real skepticism.
C2PA metadata on files is a different story. That is already verifiable today with existing content-credential tools.
What This Actually Changes For You
For most people: very little. But a few practical moves are worth making.
- If your school or employer has an AI policy, follow it. That was already true. This doesn't change the rule, it just makes it slightly more checkable.
- Own your process instead of hiding it. Keep drafts, notes, and version history. If a probabilistic signal ever gets pointed at you, a real work trail is a far stronger answer than arguing about detection accuracy.
- Do the editing anyway: for quality, not evasion. The last pass where you cut the padding, add your specifics, and put in the line only you could write is what makes AI-assisted work good. That it also weakens a watermark is a side effect, not the point.
- Don't rely on AI detectors, in either direction. They were unreliable before this and they still are. If you manage people or grade work, treat any detection result as a reason to have a conversation, never as evidence on its own.
The Risk Worth Naming
The genuine concern here isn't the technology. It's institutions treating a probabilistic signal as definitive proof, despite the company that built it explicitly warning against exactly that.
If you are the one making decisions with this, be the person who reads the limitations. If you are on the receiving end, know them well enough to point at them.
The Honest Summary
A watermark means Claude was probably somewhere in the pipeline. That's it. It doesn't know who you are, it doesn't say who wrote what, and it can't tell the difference between "Claude drafted this" and "Claude fixed my commas."
Which is a good reason to stop being afraid of this news, and a decent reason to keep doing the last twenty percent yourself.