Avoid Claude’s Watermark: 5 Easy Tips

The backlash to Anthropic’s new Claude watermarking feature has been swift and loud, but the real issue isn’t the technology itself. It’s what our collective reaction says about how we’ve come to depend on AI tools in the first place.

Earlier this month, Anthropic announced it would begin embedding watermarks into text generated by several of its Claude models. The move is designed to comply with Article 50 of the EU AI Act, which mandates that participating AI companies provide transparency mechanisms for AI-generated content. Within days, the internet was flooded with outrage, and just as quickly, a wave of watermark-removal tools started appearing on GitHub.

For users unhappy with the development, the fix seems straightforward: strip the watermark before taking text out of Claude. But this approach mirrors the proliferation of so-called “humanizing” plugins that have cluttered app stores, and it’s unlikely to solve anything meaningful. What we really need is a fundamental reset in how we view AI tools and why content provenance matters in the first place.

How Anthropic’s watermarking actually works

Initially, Anthropic offered few details about the watermarking mechanism or who could read it. That changed on Friday with new guidance explaining that watermarks are tied to specific word choices. Large language models generate text one word at a time, selecting each word probabilistically based on the one before it. In its explanation, Anthropic pointed to synonyms like “overcast” or “gray” in a sentence about weather, noting they effectively carry the same meaning.

These “low stakes” word selections leave a pattern in generated text that is invisible to human readers but detectable by those holding the right key. The watermarking process subtly alters how the model makes these random choices.

“Instead of using an arbitrary random number generator to pick the next word, watermarking uses the key and a few words that come before to settle what word the model should pick,” Anthropic explained. “The words that Claude picks are still random, but now, one can check the sequence of words and see if it’s consistent with the choices Claude would make if it was using the key. If it is, one can assign a probability that the text was generated by Claude.”

Why the outrage misses the point

John Gruber, co-creator of the Markdown markup language, has been one of the most vocal critics. In a recent blog post, he disputed the notion that these word choices are “low stakes” and argued the watermarking key could “corrupt the semantics” of Claude’s output. He’s far from alone in that concern.

“The exact words we choose when writing matter,” Gruber wrote. “I want any LLM I use to choose the very best, most precise words at every single decision point.”

As someone who writes for a living, I understand the argument that precise language matters. But is Claude really the arbiter of “the very best, most precise words”? That standard is inherently subjective, and I’m not convinced an LLM trained on the average of human language should be the gold standard. Wouldn’t the better solution be to make those word choices yourself, rather than outsourcing them to a product that keeps evolving beyond your control?

I also recognize how unhelpful it is to tell someone who finds writing excruciating to simply “figure it out.” AI can serve as a writing tool for non-writers much like Google Sheets functions as a rudimentary data tool for non-analysts. Outsourcing writing to an AI doesn’t appeal to me personally, but I say that without judgment. AI use or abstinence isn’t about moral superiority, and that dead-end argument only distracts from the larger conversations we should be having.

Gruber’s blog title, “Anthropic’s ‘Watermark’ Text Adulteration in Claude Is a Perversion of Writing,” rests on several assumptions worth examining: that Claude’s text generation qualifies as “writing” (a long-standing debate); that the “perversion” lies in the watermark rather than in outsourcing writing to a machine; and that watermarks, not pre-training, model updates, or shifting safety standards, are the first major factor to skew Claude’s output.

But one line from Gruber stood out most: “The idea that anything other than my needs should factor into the generation of text for me is patently offensive.”

Here’s the uncomfortable truth: Claude has always been a product owned by a tech company. It was never optimized solely for any individual user. Consumer-facing AI might be unprecedented, but a tech company behaving like a tech company under capitalism and global policy is not.

We owe ourselves a clearer view of this reality. Believing that Claude or any similar product functions as your personal assistant first is a logical fallacy, no matter how successfully it’s been marketed to you. All the time you’ve spent training Claude on your preferences doesn’t change where its loyalty lies. Anthropic’s decision to adapt, however haphazardly, to real policy like the EU AI Act is just a reminder of that.

AI labs are companies with their own whims and plans. This won’t be the last time one changes its product on you. What we can preserve is our agency. Instead of expecting unrealistic consistency or priority treatment, invest in the writing you want your name on. Be the copy you want to see in the world.

The case for AI transparency

Policy isn’t always about improving product quality. Sometimes it’s about addressing what a company hasn’t done, or hasn’t done enough of. That’s the intent behind Article 5 of the EU AI Act.

“It’s unacceptable for a tool to sacrifice an iota of clarity, coherence, meaning, quality, etc. for the purpose of embedding hidden clues within the text to suggest its provenance,” Gruber wrote.

Yes, AI tools have market incentives to deliver high-quality responses. But provenance, despite being an imperfect science, matters deeply. Global trust in news hit a new low this year as more readers turn to chatbots for information. Misinformation is shaping elections, and human relationships with chatbots, especially among children, are eroding our understanding of where information comes from and whether to trust it. Google now lets users toggle off Gemini’s sparkle watermark on AI-generated images and video, though hidden markers like SynthID and C2PA credentials remain.

Anthropic’s watermarks are still a work in progress. While detectors will eventually be able to read them, the public doesn’t yet have an accessible tool like Google’s SynthID Detector, on which Anthropic’s watermarks are based. AI detectors themselves are notoriously unreliable. It’s unclear how much these new watermarks will help with the information pipeline, academic integrity, or related concerns. But provenance is a field worth investing in, and if it affects your experience of an AI tool, that might be a good reason to reconsider how you use it.

Additional reasons to write yourself

Some criticisms of the watermark controversy are entirely fair. Sara Simeone, founder of NoCodeLab.ai, made a pointed observation on LinkedIn: “These labs scraped our content for years to build their wealth, without consent. Then they built one of the most ingenious revenue models ever: we pay them for regurgitated versions of what we already produced. Now we don’t only pay. We also have to ‘attribute our own IP to them’.”

There’s an undeniable hypocrisy in watermarking the product of pirated books, especially after settling lawsuits with authors over that very practice. OpenAI similarly destroyed its trove of book data. And that doesn’t account for the countless amounts of internet content AI companies have scraped, with creators uncompensated, to feed their data-hungry models.

(Disclosure: Ziff Davis, ZDNET’s parent company, filed an April 2025 lawsuit against OpenAI, alleging it infringed Ziff Davis copyrights in training and operating its AI systems.)

Then there’s the question of where watermarking ends. John McCarthy, opinion editor at The Drum, raised a valid point: “I’m even working on something that’ll help me sort and respond to pitches. Should that be watermarked too? We’re using AI to generate captions on our videos, and transcribe our interviews too. Watermark that? We’re technically recording those on electrical devices. Is that AI? I have a soft AI filter on my zoom call to reduce my ugliness. Better watermark that too?”

He’s right that not all editorial-adjacent work done with an AI tool should be flagged as AI-generated. Anthropic hasn’t clarified how it will handle these distinctions, and its initial announcement admitted that watermarks may or may not attach to smaller-scale edits made with Claude.

Beneath both posts lies a deeper reaction to Anthropic and AI labs generally: getting the economy hooked on tools that automate your work, then placing a tracker on that work in ways that could jeopardize it, feels like betrayal. But you can’t be betrayed by a company that was never beholden to you in the first place.