Anthropic Watermarks AI Use

If you have been using AI in your writing, even if it’s just for proofreading, you might want to pay attention to this article.

As of August 2026, Anthropic’s Claude, which is one of the AI models many authors use, started adding an invisible watermark to every single piece of text it generates.

And it’s not optional.

You can’t turn it off.

It’s like a bad rash.

But before you start wailing at Anthropic and cancelling your subscription, you should probably know that 200 companies have signed up for this, including Google and Microsoft.

Watch on YouTube

The culprit? The European Union (because of course it is).

The EU made a law that requires AI generated content to be identifiable. So, these companies are just complying with laws that the EU put in place, a lot like they have done with GDPR and the new accessibility laws for digital products.

Google is already participating and Open AI is expected to follow.

So, using a different AI isn’t going to solve your issue.

I know what you’re thinking though. I bet you are thinking you can strip out the AI mark by pasting it into another AI or platform. In fact, I have seen some comments regarding this. And it’s kind of laughable.

The reason is that Anthropic hasn’t revealed how to detect the watermark, so all of these ideas of how to game the system are just guesses at best. No one really knows if it works.

But let’s back up a second. Because there is something important you should probably know. There is absolutely no difference between “100% AI written” and “I used Claude to proof my work” when it comes to this mark.

How does it work?

Well, the watermark is built directly into the model itself. It works by subtly influencing which words the AI chooses as it writes. And you can’t see it and it won’t get eliminated by editing.

This mark can survive you copying and pasting the text somewhere else.

And for authors, you need to know that 100% AI slop and AI proofreading are all clumped together as “AI use”. Which sucks honestly. Because not everyone is gaming the system and pumping out AI content like an illiterate grifter.

Claude admits this in their documentation, but before you panic, it’s also important to know that a watermark is just a notification that an AI tool was used. It’s not a verdict.

But some people are going to throw hands if anyone even uses it at all and this is where authors are finding themselves in defense mode.

So how do you protect your work?

Here’s 4 ways you can protect yourself from accusations.

1. Keep your draft history.

   Whether youre writing in Google docs or Scrivener, if you save your version history, this is evidence of the draft before AI touched it. It’s your proof of authorship. That paper trail is valuable and better than a watermark.

2. Use AI for feedback not full rewrites when it counts.

   Instead of asking AI to rewrite a paragraph for you, ask instead what’s not working. Then go fix it yourself. You know, manually…how we used to write before AI did it for us. And that’s better anyway especially since your voice as a writer matters. Trust me, no one wants to read an AI polished manuscript that sounds like everyone else.

3. Understand what the mark actually proves and what it doesn’t prove.

   If it ever comes up on a platform such as Substack or with a publisher, you can honestly say, “Yes, I used AI as a proofreading pass, but everything else was written by me. It’s the same thing I’d do with a human editor.” And that’s the truth.

4. Don’t use AI tools to remove your AI usage.

   These AI watermark remover tools don’t actually remove AI. Plus, you are putting your work through another AI tool which means you are using AI even more than suspected! Don’t risk ruining your work.

All of this is kind of wild if you ask me. The policing of AI. It’s crazy. Can you imagine if we did this with calculator use, Google search use, or phoning a friend? All jokes aside, if AI is “just a tool” why are we punishing everyone that uses it?

A Study on AI Usage

I first learned about the Anthropic story from Dale L. Roberts on his podcast where he also talked about a study from Cambridge University. The study was done with 18 to 81 years olds and it included short stories, some written by humans, published in real literary journals, and some generated by AI. They wanted to see if anyone could tell the difference.

Readers couldn’t tell which stories were AI and which were human. And even more noteworthy, the stories that were written by AI were rated higher because these people believed that they were human written.

In the context of everything I’ve mentioned so far, it’s interesting that we are trying to label and tag AI use as if it’s use isn’t a good thing. Readers can’t tell the difference, but let’s watermark it and shame everyone. Sounds like a great plan.

Substack AI Detection

And this isn’t just talk.

Substack rolled out an AI detection tool. I’m sure you know more than me since you use this platform a lot more than I do. But the powers that be at Substack introduced a score showing how much of your post is using AI. There are some options to turn it off, or you can write to them and ask for it to be removed all together. But the response has been pretty loud.

Some writers are calling it a witch hunt. Some are sharing stories of getting flagged incorrectly, ruining their reputations. It’s a shit show.

You can read this NPR interview here.

But my main point is this: All of the policing that is going on with watermarks, detectors, and labels is a little absurd when you look at that study.

Shouldn’t we be happy that a story is well written and we enjoyed it? Why are we now putting content that was once valuable and helpful, or even entertaining, under a microscope?

I mean, on one hand, I absolutely understand the trash talking about AI.

Sam Altman recently was quoted as saying, “I cannot imagine figuring out how to raise a newborn without ChatGPT”.

The amount of trust issues I have with these people is through the roof. Here’s an idea Sam. Maybe don’t reproduce? Poor kid.

Anyway, wondering what your thoughts are on all of this. Leave a comment on the Substack post & let's discuss.

Links:
🔗 Dale L. Roberts' channel (go say thanks for the study tip)

🔗 Inc - Anthropic's New AI Watermark Sparks Backlash from Claude Subscribers

🔗 EU Article 50 of the AI Act