Tuesday, August 11. Anthropic announced that Claude, the AI behind a feature in my writing app, will embed an invisible watermark in the text it generates. Wednesday, the update carrying my disclosure went to Apple. Between the two, I wrote a paragraph, and the reviewer cut its most accurate sentence.
The sentence was true. That was the problem.
What actually changed
A watermark here is not a logo. It is a statistical pattern in the word choices, invisible to a reader, that a detector can later find. Anthropic says the mark travels with copied text, may survive some editing, and will eventually be detectable by third parties. It applies everywhere Claude is used, including the plumbing that lets an app like mine send text to Claude and get text back. Newer models carry it from launch. Older ones are being retrofitted, on a timeline nobody has published.
One line in their notice mattered more than the rest. To anyone who builds Claude into a product: you should independently assess what the law requires of you.
Which is a polite way of saying: your product, your sentence.
The app that had to say something
Cloister Ink is a plain-text writing app for people who write long things: theses, manuscripts, that kind of work. It has a feature called Ask. You select a passage, ask a question or request a rewrite, and the text goes to Claude using the writer's own account. The app never sees a key it did not have to.
So a novelist could paste an Ask rewrite into a manuscript. A doctoral student could paste one into a chapter. Some publishers, agents, and contests have AI-disclosure policies. That user deserves to know a mark might now ride along.
Nothing in the app broke. Nothing in the app changed. The world around it changed, and the help document had a hole in it.
The sentence I deleted
I drafted the disclosure, then handed it to a review agent whose whole job is to find claims that will age badly. Its verdict: publish, with fixes. Then it wrote a suggested sentence of its own, and I cut that too.
The cut sentence said, more or less: the models Ask uses today are not watermarked yet.
Accurate the moment it was written. Here is why it could not ship.
Ask's list of available models is not hard-coded. It refreshes live from the writer's Anthropic account. The day Anthropic finishes retrofitting an older model, or the writer's account gains a new one, that sentence goes false. No app update. No release notes. No one at my desk finds out. The lie would sit in the help document, correct-looking, until a reader caught it.
I have started calling these shelf-life sentences: statements that are true on the day of writing and depend on someone remembering to come back. Nobody comes back.
So the shipped paragraph asserts no current state at all. It describes the mechanism. Here is what a Cloister Ink user reads today, verbatim:
Anthropic has announced that Claude models embed an invisible, machine-readable watermark in generated text, applied worldwide (not only in Europe) to meet the EU AI Act's transparency rules. The mark can travel with text that is copied and pasted, may survive some editing, and Anthropic plans to let third parties detect it. Newer models carry it from launch, Anthropic is extending it to earlier models, and Ask's model list refreshes from your account, so coverage can change without a Cloister Ink update. A detected mark shows Claude processed the text, not who wrote it. If a publisher, agent, or contest has an AI-disclosure policy, treat any passage drafted with Ask's help as AI-assisted, whether or not it carries a mark.
Every clause there stays true whichever way the rollout goes. The last sentence is the only advice in it, and it does not depend on the watermark existing.
Where it went: the help document, the FAQ, and one line on the screen where a writer enters their key. Where it did not go: the App Store description, because a third party's roadmap in an immutable listing is the highest-drift sentence I could write. And not the privacy policy, because the watermark is a property of what Anthropic sends back, not of how my app handles data. Keep the policy about the policy.
If a fact can change without you, describe the mechanism, not the state.
The rule
Somewhere in your product, or your team's wiki, or your own notes, there is a sentence that is true today and that nobody owns keeping true. It was written by a careful person on a good day. It is expiring quietly.
What sentence is it? Hit reply. I read everything.
One more thing, and it belongs in this issue more than any other. Wisdom Nexus is written the way I keep telling you to work: with AI in the loop, out in the open. So the words you are reading may well carry the very mark this issue is about, and I am at peace with that. This was never a newsletter about working with AI from a safe distance. It is the practice itself, done in front of you. Describe the mechanism, not the state. That one holds for me too.
Christopher
Wisdom Nexus is written the way it is lived, with AI in the loop and a person at the desk who cares how it lands. This is where the theory meets the practice, shared with you in the open.
P.S. The scoreboard: still mostly me. The machine that moves it got better this week, and one moment stands out. I asked it to read Apple's records directly instead of trusting my notes, and it corrected my own log twice: two things I had filed as "not yet done" were done weeks ago. Same lesson as issue three, and I still needed it again. The message is never the record. Go look at the record.