08/11/2026
This is more of a LinkedIn post (and we put it there too!) but this is an important topic we wanted to share here as well.
What does everyone think of Anthropic's announcement regarding the watermarking of content that passes through Claude?
For those of you that haven't seen the announcement, here's a summary. I'm really stuck on the 3rd item below.
1) ๐๐ฅ๐๐ฎ๐๐ ๐ง๐จ๐ฐ ๐ฅ๐๐๐ฏ๐๐ฌ ๐๐ง ๐ข๐ง๐ฏ๐ข๐ฌ๐ข๐๐ฅ๐ "๐๐ข๐ง๐ ๐๐ซ๐ฉ๐ซ๐ข๐ง๐ญ" ๐ข๐ง ๐ญ๐ก๐ ๐ญ๐๐ฑ๐ญ ๐ข๐ญ ๐ ๐๐ง๐๐ซ๐๐ญ๐๐ฌ. Anthropic (the company behind Claude) announced this today, August 11, 2026. It doesn't change how the text reads, you'd never notice it by reading the text or visually.
2) ๐๐ญ ๐๐ฉ๐ฉ๐ฅ๐ข๐๐ฌ ๐ฐ๐จ๐ซ๐ฅ๐๐ฐ๐ข๐๐, ๐ง๐จ๐ญ ๐ฃ๐ฎ๐ฌ๐ญ ๐ข๐ง ๐๐ฎ๐ซ๐จ๐ฉ๐. The trigger was an EU transparency law, but Anthropic is rolling the marking out to all Claude output globally, across the chat app, API, and other Claude products.
3) ๐๐ญ ๐๐จ๐ฏ๐๐ซ๐ฌ ๐๐๐ข๐ญ๐ข๐ง๐ , ๐ง๐จ๐ญ ๐ฃ๐ฎ๐ฌ๐ญ ๐จ๐ซ๐ข๐ ๐ข๐ง๐๐ฅ ๐ฐ๐ซ๐ข๐ญ๐ข๐ง๐ . If someone writes their own document and only uses Claude to polish, translate, or summarize it, the output can still carry the mark, the mark tracks "this passed through Claude," not "Claude authored this from scratch."
4) ๐๐ก๐ ๐ฆ๐๐ซ๐ค ๐ฌ๐ฎ๐ซ๐ฏ๐ข๐ฏ๐๐ฌ ๐๐จ๐ฉ๐ฒ-๐ฉ๐๐ฌ๐ญ๐ ๐๐ง๐ ๐ฌ๐จ๐ฆ๐ ๐ฅ๐ข๐ ๐ก๐ญ ๐๐๐ข๐ญ๐ข๐ง๐ , but heavy rewriting, paraphrasing, or translation can break it. This is based on the assumption that Anthropic's "statistical mark woven in to generation" is similar to, or perhaps even based on Google's SynthID, which uses generation "nudging" to create a detectable pattern (based on the statistical likelihood of the next word appearing).
5) ๐๐ญ'๐ฌ ๐ ๐ฌ๐ข๐ ๐ง๐๐ฅ, ๐ง๐จ๐ญ ๐ฉ๐ซ๐จ๐จ๐. A detected mark means the text may have passed through Claude but it doesn't confirm authorship, and its absence doesn't prove text is human-written either.
6) ๐๐ง๐ฅ๐ฒ ๐ง๐๐ฐ๐๐ซ ๐๐ฅ๐๐ฎ๐๐ ๐ฆ๐จ๐๐๐ฅ๐ฌ (๐ฅ๐๐ฎ๐ง๐๐ก๐๐ ๐๐ฎ๐ ๐, ๐๐๐๐ ๐จ๐ซ ๐ฅ๐๐ญ๐๐ซ) ๐๐ซ๐ ๐ฆ๐๐ซ๐ค๐๐ ๐ฌ๐จ ๐๐๐ซ.
7) ๐๐ฅ๐๐ฎ๐๐ ๐ข๐ฌ๐ง'๐ญ ๐ฎ๐ง๐ข๐ช๐ฎ๐ ๐ก๐๐ซ๐. Google's Gemini has had a similar text watermark (called SynthID) running at scale for a while. OpenAI's ChatGPT doesn't watermark text yet, but given the EU mandate they are likely working on it. ChatGPT does watermark generated images.
8) ๐๐ก๐ ๐๐๐ญ๐๐๐ญ๐ข๐จ๐ง ๐ญ๐จ๐จ๐ฅ ๐ข๐ฌ๐ง'๐ญ ๐ฉ๐ฎ๐๐ฅ๐ข๐ ๐ฒ๐๐ญ. Anthropic says it's coming, but right now there's no way to check whether a given piece of text carries the mark. Even Claude doesn't know how to do it yet.
The real questions:
-What does this mean for your business and its use of Claude?
-Does this now start pushing businesses towards using open weight models they have more control over?
-Will spam filters start looking for this mark as a measure of where your email lands?
One thing is for certain, if you're building your business around today's AI capabilities, flexibility will be key. The landscape is changing at a speed that's hard to keep up with and your business will need to move at the same speed. Need help figuring this out? We're here to help your business adapt to this rapidly changing world.