Anthropic is Introducing Text Watermarking… But How? - DTNS 5329

**Jason Howell** (0:05)
This is the Daily Tech News for Tuesday, August 11th, 2026 We tell you what you need to know, give you the important context, and help each other understand.

**Tom Merritt** (0:13)
Yeah, here's something I want to understand. Today, Anthropic announced that it's rolling out text watermarking.

**Jason Howell** (0:20)
Uh, uh?

**Tom Merritt** (0:22)
How do you watermark a text without changing it? And they're also doing C2PA.

**Jason Howell** (0:27)
That's a real good question.

**Tom Merritt** (0:28)
Maybe you made them do it, but all of us get to benefit.

**Jason Howell** (0:33)
We're going to talk all about it. Coming up here, I'm Jason Howell.

**Tom Merritt** (0:35)
I'm Tom Merritt.

**Jason Howell** (0:36)
Let's start with what you need to know with a big story.
So, yes, Tom, as you eluded to, uh, Anthropic shared in a help document, uh, that its models will carry embedded text watermarks, also C2PA signed file metadata. This is all in compliance with the EU AI Act, but that might make you think, okay, so they're going to get it, the rest of the country, the rest of the world will not.
Uh, but that is not true. Apparently, the watermarking going to be applied globally for all users. This is one of those cases where that actually happens. Uh, the watermarking covers the Cloud API, Cloud, Cloud Code, Cloud Cowork, Tag, third party hosts. So if you get your cloud through something like AWS or Google Cloud, um, it's going to survive through there as well, I guess, because it's, it's text. Uh, yeah, I don't know how this is going to work, but Anthropic says that the text mark survives copy paste actions, uh, and some editing without changing the meaning or changing the quality of the text that it's attached to. It's, man, it's so puzzling. Well, Anthropic also says it'll support its users in detecting Cloud's own watermarks with details on how that's done, coming sometime in the future. So we don't even know quite yet how you would even detect for this, but Anthropic is going to have a tool for you. Anthropic says the watermarking will come to new models, of course, but also they're working on integrating it into earlier models that were released prior to this August 2nd deadline that was required by the AI Act. So yeah, text watermarking. I mean, I'm sure text watermarking is a thing before this, right?

**Tom Merritt** (2:21)
It is a thing. In fact, Apple, more famously than most, used it to watermark some of its internal documents, which is how they were able to track a leak, because they can find the text watermark.
As I understand it, and I haven't spent my entire life studying it, but as I understand it, text watermarking takes a few different forms. It usually has something to do with word frequency or patterns. So in the English language, we have so many synonyms, I imagine it might be a little easier, but you can use synonyms or you use certain phrases, and then you can keep the meaning the same while changing the words to make sure that later you can do a text analysis and go, ah, that was made by Claude, because that's Claude's way of using those words in that manner. Similar to how you can actually identify the writer of something based on their word choice and their word usage, right? And watermarking is kind of like that.
I guess that doesn't change the meaning, but it actually as someone who wants to use these tools for like just grammar, you know, and per word, I don't want it changing things to watermark them. So like, yes, just make things bold, but don't change any of my words. And if it's trying to watermark, then it might have to change words. So I guess I'm not trying to be like, and therefore I'll never use Claude, but I'm like, I need to know a little more about how their kind of text watermarking works.

**Jason Howell** (3:53)
Right, and when is it applied? Is it like, here's my paragraph, make sure that I didn't miss any commas, and it inserts a comma and it's like, all right, text watermark in there, we had something to do with this.

**Tom Merritt** (4:04)
Yeah, yeah, I'm guessing it doesn't work like that. There are some other text watermarking that use invisible characters. You know, and that's when they say like, it'll survive most cut and paste. I'm like, oh, okay, so maybe they're just, maybe they're just putting unreadable characters in there. And then when you copy paste them, that persists. I wouldn't mind that as much because it's not messing with my copy then.