This archived newsletter issue covering the impacts of Anthropic’s Claude watermarks on freelance writers was originally sent to subscribers on August 19, 2026.
If you haven’t heard yet, Anthropic (the AI company behind Claude) is now “watermarking” text processed by Claude.
This move came in response to the EU’s AI Act.
And look… I’m all for disclosure and transparency. But the way this is being handled is misguided at best, and malicious at worst.
What it isn’t is good news for writers.
What is AI Watermarking?
Claude watermarking text it processes isn’t quite like an image watermark you might be more used to.
There aren’t any hidden pixels or hidden words or letters.
Instead, the watermarking is more programmatic and probabilistic.
Basically, the watermark is in the word choice Claude uses.
The probability of it choosing certain words and phrases will be altered in a way that makes the pattern of those choices a machine-readable sign saying “Claude was here.”
Why This Kind of AI Watermarking Misses the… Mark(?)
In theory, this might not sound bad. It doesn’t visually affect the text, and it discloses Claude’s involvement in its creation.
Or does it?
The biggest issue I have with this (and why I cancelled my paid Claude account until they resolve their issues) is that this type of watermarking will be abused.
100%. I have no doubt about this.
We’re already seeing AI witch hunts.
AI detectors have never been accurate.
Yet they’re used to accuse students of cheating.
Freelancers have lost clients because an AI detector said their own work was likely generated by an AI tool.
An author lost a multi-million-dollar book deal just on accusations of AI use.
Then you have issues on social media and elsewhere…
Everyone thinks they’re an expert at identifying AI-generated content. (Hint: Nope.)
They’re happy to accuse anyone saying anything they don’t like or disagree with of posting “AI slop” (a term that immediately tells me not to take someone seriously).
Will AI watermarking solve these issues?
No. But they might make them worse.
There’s a reason Anthropic used CYA (“cover your ass”) language when describing this AI watermarking.
It doesn’t mean “Claude wrote this,” they say. It means, and I’m paraphrasing, “Claude might have processed this.”
Well, what the ever-loving …?
They know very well how detection of this watermark (which isn’t even available yet) will be used.
And tin-foil hat time or not, I think they know exactly what they’re doing.
You see, the EU’s AI Act has explicit exceptions regarding things like editing text. AI grammar checkers and the like are fine, for example, and they don’t require these kinds of watermarks.
Anthropic decided to watermark these “light touches” anyway.
So if you use Claude to review something you wrote entirely on your own to help you refine structure, edit your grammar (maybe you’re writing outside of your first language), etc… you might get that same “processed by AI” stamp on your work as someone who outright publishes one-prompt AI-generated content.
And do you genuinely believe average readers, once they’re able to find these watermarks, are going to understand that difference, or even care?
I don’t.
I see this as feeding the existing AI witch hunts, and that’s only going to hurt writers.
After all, these tools are trained on professionally-written work. It’s why it’s not surprising when a professional writer’s work is flagged by early-generation AI detectors.
It’s also why we’re sometimes better at spotting it.
We know how em-dashes are meant to be used, so we can spot when they’re used incorrectly or too often.
The same is true when AI tools use copywriting tactics in content they generate. We can spot when they’re used incorrectly or in the wrong types of content.
Maybe I’m wrong, but…
I don’t see this as an oversight.
I think it’s more strategic than that.
Here’s why:
- Increased false accusations would cause outrage, possibly slowing down similar legislation elsewhere.
- It blurs the line of authorship. Technically not claiming “Claude wrote this” doesn’t mean it won’t be read that way. What if it’s eventually considered in copyright applications? Or just people thinking an AI watermark means it’s AI-written, therefore not copyrightable, and they can steal it and republish it as they please? It forces a new burden of proof on actual writers.
- When authorship can be that easily questioned, and copyright eligibility as well, that benefits AI companies. It’s already been decided by the courts that AI training is fair use, but they have to acquire the material they train on legally, and that doesn’t apply to AI outputs that can still be infringement. If copyright protection is in any way weakened by generic AI “processing” claims, that’s a workaround they could try to exploit.
In other words, I think Anthropic is well aware of the confusion this kind of unnecessarily-broad AI watermark will cause. And I think there are strategic benefits to both the company and industry if that happens.
There are other important considerations.
That confusion, intentional or not, isn’t the only problem with this type of AI watermark.
For example, users paying for AI tools expect them to give the best answers for their task, project, or prompt.
Yet you can’t optimize a response in all cases when you also have to stick to these probability rules for the watermarking. Shorter texts could see clunkier synonyms used because there’s less overall output to work with for example.
And here’s an even bigger issue:
Bad actors aren’t the ones who will be punished.
First, bad actors trying to pass off AI-generated content as their own won’t stop.
They’ll switch tools. And when watermarks are implemented there (hopefully more responsibly), they’ll just move to local open-source models, which keep getting better.
The ones targeted will be:
- writers who can’t afford legal battles over contracts lost to false accusations (easy way out for clients or publishers);
- groups already targeted for daring to exist and write about issues important to them (LGBTQ+ writers, POC, disabled and neurodivergent writers, etc.);
- younger and newer writers who don’t have years of published clips, books, and such to prove their “voice” is their own when people question them.
This will continue to cost people livelihoods. It will empower digital mob mentality. And it will be used to, quite literally, de-humanize writers.
So, I get it. We need more transparency.
Personally, I love sharing the ways I use AI in my business (not generating content), and I enjoy hearing how others use it as well.
This isn’t the solution though, or at least not the way Anthropic chose to handle it.
On the legislative side, we need more accountability for the companies and individuals using these tools deceptively (advertisements with fake “influencer” videos held accountable for false advertising and FTC disclosure violations, broader “right of publicity” rules that protect people’s actual images from being used without their consent via AI models, etc.).
Until then, writers are at an increased risk of being accused of not doing the very work they pour their hearts into.
I don’t know about you, but I don’t see that as acceptable.
Would you like to receive content like this in your inbox? Get subscribers-only content, and early content before it’s archived on the website, by subscribing to the Freelance Writing Pros newsletter.