◉CLAUDE LABJP
●CC 2.1.325 — No new release over the weekend; 2.1.325 (11/13) is still the latest●BILLING — 2.1.325 fixes background sessions on a claude.ai login being billed to an API key from the shell●SONNET 4.5 — Retires on the Claude API on 11/30, 14 days left. Check saved model settings too●403 — Users report a Cloudflare 403 page mid-session that still burns tokens●MANAGED — Since 11/12 you can request self-hosted Claude Managed Agents that run entirely on your own infrastructure●NEW — When Browser Extension Is Not Connected Won't Go Away●CC 2.1.325 — No new release over the weekend; 2.1.325 (11/13) is still the latest●BILLING — 2.1.325 fixes background sessions on a claude.ai login being billed to an API key from the shell●SONNET 4.5 — Retires on the Claude API on 11/30, 14 days left. Check saved model settings too●403 — Users report a Cloudflare 403 page mid-session that still burns tokens●MANAGED — Since 11/12 you can request self-hosted Claude Managed Agents that run entirely on your own infrastructure●NEW — When Browser Extension Is Not Connected Won't Go Away
Articles/Claude.ai
◉ Claude.ai/2026-08-18Advanced

What to Settle Before Watermarked Claude Text Ships Inside Your App

Claude models released on or after August 2, 2026 embed watermarks in generated text and attach C2PA-signed provenance metadata to files. Here is how I redrew the line for shipped copy, plus a measured look at where image metadata disappears and a submission gate that keeps provenance reproducible.

Claude53watermarkingC2PAprovenanceindie development26App Store9compliance

✦ Premium Article

I was lining up store listings for four apps across languages when I stopped and stared at the folder for a couple of minutes. The Japanese description was mine, written by hand. The English one was that same Japanese text put through Claude. Korean and German came from the same route. On disk they all looked identical: text I had prepared.

Then the marking news landed. Claude models released on or after August 2, 2026 weave a machine-readable watermark into generated text. The mark lives at the model level, travels through copy and paste, and may survive a certain amount of editing. Which means the English description was very likely carrying a signal that it had passed through Claude, whether or not I intended it to.

The awkward part is what that signal actually means. It says "Claude processed this," not "Claude wrote this." Translation counts. Proofreading counts. Summarizing counts. As an indie developer I had a line I was comfortable with — I decide the substance, tools help me deliver it — and this single distinction forced me to redraw it.

The two marks fail in opposite directions

Anthropic ships two mechanisms with genuinely different properties. One is hard to remove. The other comes off almost by accident.

PropertyText watermarkFile provenance metadata
Where it livesInside the model output itselfAttached to the file (C2PA signature)
Applies toAll generated textSupported types such as .svg, .png, .jpg
Copy and pasteTravels with the textNot applicable
Survives editingOften, to a degreeFrequently lost on re-save
Default failure modeSticks around unnoticedFalls off when you wanted it
Short fragmentsToo little signal to detectLost through format conversion

That asymmetry drives everything downstream. Text drifts toward shipping with a mark you never thought about. Images drift toward losing a mark you may have wanted to keep. Leave both to their defaults and you end up somewhere other than where you intended.

Anthropic is explicit about the limits: a detected mark is a signal that content may have been processed by Claude, not a full account of its provenance, and the absence of a mark proves nothing either. Their write-up is worth reading in full at How Claude marks AI-generated content.

Redraw the line around "went through," not "was written by"

My previous rule was about authorship. Anything that counts as the work itself, I make by hand. Anything that exists to deliver or promote it, I am happy to hand to a tool. Clean enough as categories go.

Watermarking does not care about authorship. It cares about processing. Feed your own Japanese to Claude for an English translation and the English is generated text. Feed your own English to Claude for a copy edit and what comes back is generated text. You wrote the substance; the artifact that ships is model output.

This was the part that ran against my expectations. People who are careful about not generating content are the ones most likely to mark copy without noticing, because translation and proofreading feel like touching up your own writing rather than producing something new.

So I swapped the test. Not "who wrote this" but "did this go through Claude." In practice it fits on one line:

Did this exact string ever come back as a Claude response?

If you took the return value as-is, it went through, even when the input was yours. If you read the response and then rewrote the line in your own words, it did not. That version of the question is easy to answer honestly and easy to verify later.

✦

Thank you for reading this far.

Continue Reading

What follows includes implementation code, benchmarks, and practical content we hope you'll find useful. This site runs without ads — server and development costs are supported entirely by members like you. If it's been helpful, we'd be truly grateful for your support.

WHAT YOU'LL LEARN
✦You will be able to decide which shipped strings may pass through Claude using a test based on whether text went through the model, not on whether the model wrote it
✦You will know exactly which export step strips provenance metadata from your screenshots, and how to verify survival before you submit
✦You will be able to prevent the situation where copy ships without any record of who wrote it, using a single pre-submission gate instead of memory
Secure payment via Stripe · Cancel anytime
✦

Unlock This Article

Get full access to the rest of this article. Buy once, read anytime. This site is ad-free — your support goes directly toward keeping it running.

or
Unlock all articles with Membership →
Share

Thank You for Reading

Claude Lab is ad-free, supported entirely by members like you. We publish practical guides daily with implementation code, benchmarks, and production-ready patterns. If you've found it useful, we'd love to have you on board.

  • ✦Copy-paste ready implementation code
  • ✦New advanced guides published daily
  • ✦$5/mo or $10 for lifetime access
View Membership →

Related Articles

◉ Claude.ai2026-10-19
Claude Said the File Was Created, but the Folder Was Empty — A Read-Back Check After Every Write
When a file-creation call reports success but nothing exists on disk, here is the order I use to narrow down the cause, plus a small script that reads back every write.
◉ Claude.ai2026-10-16
I Stopped Asking Claude to 'Cite Sources' — Three Ways I Ask Instead
Three ways I check Claude's answers before they go into a published text, instead of asking for sources: leave room to say 'I don't know', pull verbatim quotes first, and have each claim re-checked against the source after drafting. Plus what still needs a human.
◉ Claude.ai2026-10-15
Let Claude Draft the Quote's Wording, Keep the Numbers in Your Own File — A Small Source of Truth and a Check Script
How a freelancer can have Claude polish the wording of quotes and invoices while the amounts stay in a JSON file and get verified by a short Python script.