I wrote this with AI.

The core problem: The EU AI Act now requires AI providers to mark AI-generated text, and Anthropic, OpenAI, Google, Meta, and Microsoft have all signed the Code of Practice that implements it. But the act regulates the wrong thing. A watermark verifies who or what arranged the words. But what readers need to know is authorship: who developed the idea, whether it’s true, and who stands behind it. The watermark will flag honest, high-quality AI-assisted work while bad actors strip it out with a rewrite. The EU is regulating the typing layer of writing instead of the thinking layer.

Let me ask you a question: how many cookie banners have you actually read? You’ve clicked “Accept” thousands of times, and you’ve read the text behind the button… approximately never. That’s the legacy of the EU’s last great transparency project: a rule designed to inform everyone about cookies that instead trained the entire planet to click past a legal notice without looking at it.

The EU just ran the same play again. But this time, you don’t get a choice. The watermarking lives inside the words generated by the AI, is invisible to human readers, and can only be detected by someone (or something) with a key.

All future Claude models released after August 2 will watermark the text they generate. Anthropic signed the EU’s Code of Practice on Transparency of AI-Generated Content in July, alongside roughly 190 other signatories, including OpenAI, Google, Meta, and Microsoft. The mark carries no information about who you are, travels when the text is copied, and may survive light editing.

I have skin in this game, because I write with Claude and I say so openly. So let me tell you what this watermark actually measures, and why that’s the wrong thing.

Your company is about to be scanned

Your company has skin in it too, even if nobody on your board has ever typed a prompt. A telco is one of the biggest content factories on earth: customer emails, customer support responses, outage notices, executive bylines, earnings scripts, RFP responses, board memos. As of this month, everything your teams produce with a frontier model will carry a mark, and there is no procurement path around it short of running your own models—and nobody is moving their customer comms onto a self-hosted model to dodge a watermark. Anthropic has said a watermark detection API is coming. Journalists, regulators, competitors, and angry customers will soon be able to scan anything you publish. “Telco’s breach apology was written by AI” is a headline waiting for a slow news day, and the damage will come from looking like you hid it, because almost nobody actually cares that you used the tool.

You didn’t vote for this

Anthropic is applying the watermark globally because it doesn’t yet have a durable way to scope it by region. The EU wrote a rule for 450 million Europeans, and eight billion people are getting it by default.

Telco has been through this before. GDPR compliance became the global privacy baseline for the same reason: maintaining two regimes costs more than surrendering to one. The pattern even has a name: the Brussels Effect. It happens because applying the most stringent rules to everything is always the easiest, cheapest option for the organizations who have to comply.

And look, the EU isn’t wrong that AI-generated content is flooding the internet. Provenance signals have legitimate uses: catching deepfakes and flagging dangerous, misleading news reports. But when a transparency rule ships to the entire planet without a vote, we should ask what it actually illuminates. 

That’s where this one falls apart.

Provenance ≠ authorship

The watermark answers exactly one question: did a machine arrange these words? That’s provenance, a chain of custody for text.

Now think about what you actually want to know when you read something. Who thought this? Is it original? Is it true? Will someone put their name on it and take the heat if it’s wrong? That’s authorship. The watermark doesn’t answer any of these questions. In fact, it confuses the issue.

Consider a few examples that carry no watermark: the CEO letter ghostwritten by a comms team, the bylined op-ed drafted by an agency, the analyst report assembled by junior staff and published under a partner’s name, and the “thought leadership” piece the executive never read before it ran. Our industry and others have relied on unmarked, unauthored content for decades. Nobody in Brussels lost a minute of sleep over it. Great content is often the result of teams working together: a collection of experts, different perspectives, multiple sets of eyes. Why does that change when one of the team members is AI?

Now consider what will get marked going forward: an argument built over weeks by someone who uses AI to experiment with ideas and structure, and type it at speed. The provenance includes AI—and so what? The authorship is clear and genuine, start to finish.

The worst part is, the watermark won’t necessarily stick to the people it’s supposed to catch. A heavy rewrite or a translation pass can knock it out. Knowing that, a content farm laundering thousands of articles a day can remove it by adding one more step to its pipeline. Meanwhile, the honest writer who says “yes, I use AI” wears that badge forever. The detector catches the disclosed and misses the deceptive. It is precision-engineered to flag exactly the wrong population.

You can see how this all might become little more than a detection ritual. Browser plug-ins, platform flags, HR policies, submission checkers, may all be scanning away while the fraudsters and other criminals keep on flying under the radar. And the ritual will harden into a proxy. People will treat the watermark as a lie detector for thought: unmarked text will read as genuine thinking, and marked text will read as thinking outsourced to a machine. Both inferences fail. And a test that fails in both directions is worse than no test at all, because people will trust it.

Next comes the virtue signaling. “Human-written” may become a purity label, brandished the way food brands flaunt “no artificial ingredients”: a claim aimed at your feelings that says nothing about what’s actually in the box. LinkedIn will fill with no-AI pledges. Agencies will sell “certified human” at a premium. The claim is unverifiable, so the loudest pledges will come from operations that always outsourced their thinking anyway. The mark may have been meant to inform, but it’ll be used to perform.

Brussels will argue that the mark was never designed to judge authorship. Doesn’t matter. Regulations are judged by how they get used. So when everyone starts asking “was this AI-generated?” understand that it’s the wrong question. Instead, we should be asking “who is accountable for this?”

Define your position now

The way I see it, we all write with Claude—or ChatGPT, Gemini, Grok, Perplexity and others. We’re all saving a lot of time (and typos), and that’s not a bad thing.

But this law, and the watermarking, has created a situation where we should define our position before a detection tool decides it for us. The tempting response is an internal detection ritual: ban watermarked text in deliverables, scan the drafts, interrogate the interns. Sounds like a nightmare, right? It is, and it’ll fail on day one.

The durable response is to lean in. Announce that your company writes with AI—in your brand voice, on your own timing, before anyone runs a scan. The watermark only has power over organizations that hide something. If you disclose first, then the detection API just points at an empty room. Also, put a human owner’s name on every artifact: someone who read it, verified it, edited it, and answers for it. Ownership has consequences. Disclosure kills the gotcha, and ownership answers the questions the watermark never will.

I write with Claude

I’ll go first. Here’s how these blogs get written. I develop an argument over days of back-and-forth with Claude: drafts generated, graded and thrown out; sections cut; claims killed because they couldn’t survive a smart reader’s pushback. (This blog went through 17 revisions.) All this happens way faster than it used to, when a small team of writers helped with the first drafts. Not only does Claude type faster than us humans, but it’s so fast that I can also instantly course-correct, take out whole sections, or just start over if an idea doesn’t pan out, or if I wasn’t crystal clear the first time I asked. I can take something through multiple rounds in an hour. In fact, I’m actually MORE involved in the early drafts of my content now than I was before AI. What hasn’t changed is what’s behind it all: my point of view. Mine comes from thirty years in enterprise software and a decade of watching this industry repeat the same mistakes with new vocabulary.

I write with Claude. I’ll keep writing with Claude. Every word will carry the EU’s invisible mark, and every idea will still be mine. If a detection tool someday tells you a machine touched this text, it’s working perfectly. But no model will be able to generate a TelcoDR blog, because it’s my authentic thought.

The watermark knows what typed this. You’re reading it because of who thought it.

Recent Posts

Get my FREE insider newsletter, delivered every two weeks, with curated content to help telco execs across the globe move to the public cloud.