Detection Tells You Who Typed. It Can't Tell You Who Decided.
What Just Happened
On August 14, 2026, Anthropic explained how Claude's text watermark works. Using SynthID-Text — a method published by Google DeepMind in 2024 — all future Claude models embed a statistical pattern in their word choices: invisible to any reader, verifiable with a key. A detection API is coming. The driver is the EU AI Act, signed in July 2026, with roughly 190 signatories already on the EU Code of Practice on Transparency of AI-Generated Content.
In plain terms: text written by Claude is becoming checkable. Paste a passage into a detector, and it can tell you — with confidence that grows with length — whether a machine chose those words.
The watermark carries no identifying information. It cannot name a user, an organization, or a conversation. It answers exactly one question: did a model pick these words?
What It Can and Cannot Do
The limitations matter as much as the capability, and Anthropic states them plainly.
The watermark survives normal use but not a complete rewrite. It is sparse in text where there was nothing to choose — exact math, precise code, single-correct-answer passages. It is weak on short samples. It cannot distinguish "Claude wrote this" from "Claude heavily edited this." And it cannot do the inverse: the absence of a watermark does not prove a human wrote anything, and it says nothing about other models.
So it is a one-way signal. Presence means machine word-choice. Absence means nothing at all.
Keep that asymmetry in mind, because an entire genre of marketing is about to collide with it.
The Claim That Just Acquired an Expiration Date
There is a brand position that has been growing for two years: "100% human-made. No AI." It sells authenticity by denying the tools.
That position just became a testable claim. Not eventually — this quarter, as detection APIs roll out. One pasted paragraph is all it takes. And the first brand caught watermarked after promising "no AI" won't have a quality problem; it will have an honesty problem, which is much worse.
The inverse position — "AI-powered" as a badge — fares no better. When every company's text can carry the same watermark, having used a model distinguishes exactly no one.
Both positions made the same mistake: they made the tool the message. The tool was never the message.
Why We Signed Differently
Every platform, deliverable, and page we ship carries the same signature: Created with human intelligence.
Read it carefully — it does not claim the absence of AI. We build with AI, at machine speed, and we say so plainly. A signature that denied the tools would have collapsed at the first inspected artifact; we wrote ours knowing that day would come. It just came faster than expected.
What the signature claims is the thing no detector can test: who decides. The human chooses the scope, sets the criteria, and verifies before anything is delivered. The machine executes the repeatable. Nobody signs what they haven't reviewed. Our manifesto compresses it to five words: the human doesn't produce — the human guarantees.
A watermark is metadata. A signature is liability. One labels text; the other binds a person to a result.
The Questions That Actually Matter
Here is what gets clearer once detection is normal: clients never actually needed to know whether a machine chose the words. That question was a proxy.
The real questions have always been: Who decided this was correct? Who verified it before it reached me? Who answers if it's wrong?
None of those live in the statistics of a text. They live in method — in verification gates, in provenance, in someone with a name accepting responsibility for the output. That is what we sell, and it is the part of the work that was never automatable, because responsibility doesn't run on a GPU.
Detection tells you who typed. It can't tell you who decided — and who decided is the only thing that was ever worth paying for.
What Changes for Us
Nothing — and that is the point.
Our text may carry Claude's watermark. Coherent: we never said otherwise. If anything, the watermark becomes protection — our claim was built to survive inspection from day one, and now there is an industry-wide way to check what it never denied.
We welcome transparency rules. A market where claims are testable favors whoever sells proofs. In a world of promises, that has been our whole bet.
Created with human intelligence. Now verifiable in the only way it ever needed to be: by the method behind it.