On Aug. 11, Anthropic announced it would begin adding machine-readable watermarks to Claude’s outputs. The reaction was immediate and predictable. LinkedIn and X filled with the usual takes:
- “All AI writing is now fully traceable!”
- “This is the death knell for AI content farms!”
- “SEO is dead. Again.”
Sensing the uproar, Anthropic quickly followed up with a blog post, FAQs, and a technical demo showing that the watermark had no practical effect on output quality.
A few days later, Dario Amodei posted on X about AI’s broader crisis of trust, arguing that the public’s skepticism runs deeper than any one company’s messaging.
The technical explanations were clear. The demo was impressive. Yet public reaction remained largely negative.
In this article, I want to separate the hype from the reality and explore why what appeared to be a straightforward regulatory compliance announcement may instead become a flash point dividing the Eloi who embrace AI from the Morlocks who oppose it.
A quick history of watermarking
Craftspeople have marked their work for centuries.
In 1266, the English Parliament required bakers to use distinctive marks on their bread. By 1282, papermakers in Fabriano, Italy, were creating translucent watermarks with wire molds embedded in the paper.
The principle was simple: this is someone’s work, and the maker should be identifiable.
In the digital era, stock image libraries adopted the same idea. You’ve seen Shutterstock’s repeating patterns and Getty Images’ overlays stamped across preview images. The goal was the same: identify the original creator and discourage unauthorized use.
The EU rule Anthropic is answering
Anthropic’s decision is a direct response to Article 50(2) of the EU AI Act (Regulation 2024/1689). The provision requires providers of systems that generate synthetic text, images, audio, or video to mark those outputs in a machine-readable format so they can be detected as artificially generated or manipulated. The technical measures must be effective, interoperable, robust, and reliable, “as far as this is technically feasible.”
That final phrase carries significant weight. It’s not a precise legal standard.
To give companies a practical compliance path, the EU published a Voluntary Code of Practice on Transparency of AI-Generated Content. Most major providers (Anthropic, OpenAI, Google, Meta, Microsoft, Mistral, Cohere) signed it. xAI did not.
What ‘text watermarking’ actually means here
The term itself is causing confusion, so it’s worth being precise.
Traditional text watermarking typically relied on orthographic steganography: inserting hidden characters, zero-width spaces, or other invisible markers into finished text. These methods alter the form of the text. Once you know what to look for, they’re relatively easy to detect and remove.
Anthropic is using a different approach: statistical, or generative, watermarking.
When a language model generates text, it doesn’t always choose the single most likely next word. Instead, it samples from a range of plausible candidates. That controlled randomness helps keep the writing from becoming flat and repetitive. Statistical watermarking replaces some of that randomness with choices guided by a secret key. To the user, the output still appears natural. To the provider, the sequence of choices creates a detectable statistical signature.
Anthropic has said the method doesn’t insert hidden characters, identify individual users, or have any practical effect on output quality. A developer also released a demonstration tool based on the SynthID-Text approach. The engineering is sound.
Yet public reaction remained largely negative, even after Anthropic’s explanations.
That’s because the company answered the technical objections while largely missing the concerns that matter most to the people who use these tools every day — or who still need convincing to use them.
The real problems
1. It treats AI use itself as the problem
Imagine buying a set of kitchen knives and having the government assign someone to monitor you around the clock to make sure you don’t stab anyone. Don’t worry, they say. As long as you only use the knives to cut vegetables, you’ll be fine.
That’s the logic behind this approach.
Historically, watermarking existed to protect creators. Here, it’s meant to protect the potential victims of people who use AI.
Yes, scammers will use AI for fraud. Yes, people will be misled by synthetic content.
Those risks are real. But this policy rests on the assumption that the default use of AI is suspect, so the tool itself must bear a permanent mark.
Anyone who’s worked in SEO has seen this pattern before: white text on white backgrounds in the 1990s, paid links in the 2000s, private blog networks in the 2010s. The tactics worked for a while, then the market and the platforms adapted.
We didn’t need a special regulatory regime treating every form of content creation as potentially fraudulent. Existing fraud and consumer protection laws, along with Google’s incentive to protect the quality of its search results, were enough.
AI is a tool. It can be used well or poorly. Building the system on the assumption that users can’t be trusted isn’t a good way to earn their trust.
2. A positive detection becomes a Scarlet Letter
This is the practical issue that matters most to people doing the work.
Statistical watermarking can’t distinguish between high-value and low-value uses. If Claude performs light editing, rewriting, translation, or tone adjustment, the output can still carry a watermark. The watermark indicates the text was processed by Claude, not that Claude was the original author.
That distinction will be lost on most people. In practice, a detected watermark is likely to become a negative signal — a sign that the work is somehow less legitimate. Ironically, the people producing the lowest-value content will have the strongest incentive to strip or evade the watermark. Its absence will prove almost nothing.
The technique also isn’t especially durable. Just when we thought we were past the endless “we cracked Google’s algorithm” cycle, we’re about to start the same cat-and-mouse game again. Once reliable detectors exist, people will test how much paraphrasing, human editing, or multi-model processing it takes to weaken the signal.
3. It treats writing like a math problem to be optimized
I studied both computer science and English. When I read Anthropic’s explanations, the computer scientist in me was intrigued. The description of the sampling process was clear, and the demonstration tool was genuinely instructive.
The English major in me cringed.
Read these three sentences and see if you can spot the difference:
- Four score and seven years ago our fathers brought forth on this continent, a new nation, conceived in Liberty, and dedicated to the proposition that all men are created equal.
- Eighty-seven years ago, our forefathers established upon this continent a new nation, born in liberty and devoted to the principle that all men are created equal.
- Fourscore and seven years past, those who came before us brought into being on this continent a new nation, conceived in freedom and committed to the truth that all men are created equal.
From a narrow technical perspective, all three are grammatical, coherent, and “high quality.” From the perspective of someone who values good writing, only one is doing the work of literature. The other two are competent paraphrases.
An engineer or computer scientist might not even notice the difference. Readers will.
AI writing already has recognizable patterns: a heavy reliance on em dashes, the familiar “It’s not X, it’s Y” construction, overuse of words like “delve,” “leverage,” and “underscore” where simpler language would do, neatly balanced but empty phrasing, and a lack of specific, independently verifiable details that could only come from real experience.
Adding a statistical bias on top of those tendencies introduces another artificial constraint on the output. The stronger the required signal, the more constrained — and less human — the writing is likely to feel.
4. It applies a regional rule globally
Anthropic didn’t write the EU regulation; it’s simply responding to it. Still, the decision to apply the watermark worldwide at launch, rather than limiting it to the jurisdictions where the law applies, was deliberate and speaks volumes.
The company’s stated reason was the “lack of a durable way to scope the feature by region.” That may be technically inconvenient, but it’s hardly impossible.
Companies routinely adapt product behavior to local legal requirements. Choosing not to do so here — especially for a user base that extends well beyond the EU — suggests a surprising disconnect from its users, many of whom are sophisticated enough to switch to open-weight or non-watermarked models when they want maximum flexibility.
The deeper problem
On the surface, the past week looks like a tech company solving a technical problem to meet a regulatory requirement. To Anthropic’s credit, it moved first and was transparent about the change.
Where it went wrong was the audience it seemed to be addressing. Its explanations were clear to people who already understand how language models work. They did little to address the broader crisis of trust.
A few days after the announcement, Dario Amodei posted on X that the public’s negative view of AI is fundamentally a crisis of trust.
- “I do agree that the public has a negative view of AI (and that this is a big problem), but I don’t think it is primarily caused by me or any other AI leader warning about AI’s risks. I think it is fundamentally a crisis of trust.”
He has the diagnosis right. What’s less convincing is the cure.
He went on to argue, correctly, that glitzy marketing won’t fix the problem, and neither will simply claiming AI will cure cancer. The real solution, he suggested, is actually curing cancer.
That framing misses the point. It’s a blind spot shared by many AI executives.
AI won’t cure cancer. Humans will.
AI can surface connections, identify patterns, and accelerate parts of the work. But it’s still a tool. Behind every meaningful result is human judgment and human responsibility.
The same gap appears at a more ordinary level.
Outside of work, AI has improved my life. I’ve already shared how it helped me improve my health. I’ve also used it to plan vacations, adapt recipes, repair my car, and research my family history.
None of those uses will change the world. But they changed mine. Not because I picked the right model, but because I knew how to use it.
I’ve found the same is true for many long-time SEOs. Good SEOs know how to ask questions. We know how to challenge what a computer gives us, refine our prompts, and decide when to accept an answer and when to push back.
Most people haven’t had that experience. Their exposure to AI is largely limited to viral videos and a steady stream of horror stories: mass layoffs, data centers straining local resources, and executives accumulating fortunes that would make the old robber barons blush. With all due respect to Amodei, actually curing cancer won’t change any of that.
Talking as though the technology itself will deliver the breakthrough turns people into spectators instead of participants. Worse, some hear that message and conclude the companies quietly share Agent Smith’s view in “The Matrix”: humans are the problem, and AI is the solution.
What will close the gap is the same force that drove mainstream internet adoption in the 1990s: people discovering tangible benefits in their own lives. That happened because the early internet was built in a spirit of openness rather than control.
The internet scaled because its architects favored open protocols and worked in a culture that was skeptical of concentrated power, whether in government or corporations. Vint Cerf, Bob Kahn, Tim Berners-Lee, Jon Postel, Linus Torvalds, Richard Stallman, Paul Mockapetris, and many others still aren’t household names. Most never became multimillionaires or sought public recognition, yet their contributions to daily life are immeasurable. The political class’s greatest contribution was restraint.
Today, the major AI labs are responding to pressure by adding constraints and tightening control. Too often, the visible motivation seems to be who can produce the biggest exit. That’s a very different spirit from the one that built the early internet.
What actually matters
There’s a useful parallel here for SEOs. You’ve always been able to distinguish between using a technique to create real value and using it to game the system.
This article is a good example. I wrote it the old-fashioned way, drafting it myself and using AI only for research.
Once I had a draft, I used AI to organize, prune, and refine it. I didn’t blindly accept every suggestion. I pushed back and, in some cases, overrode it.
A good example is the H.G. Wells “The Time Machine” analogy above. AI kept urging me to expand that paragraph and explain the reference. I said no. I think enough of this audience will get it immediately. The rest of you can spend five seconds Googling it (or, better yet, check the book out from your local library).
The difference between quality work and slop isn’t whether it passes a detection tool. It’s whether people engage with it, share it, and convert. Everything else is secondary.
It’s also telling which tool I chose. I’ve been using Claude all month for real work. For this piece, I switched to Grok precisely because it doesn’t fingerprint its output.
Part of that decision was rational. Part was emotional. Companies ignore that mix at their own risk.
Read more at
