It was Tuesday morning, two days before this piece first went out. I was staring at my half-empty coffee mug at 9am, thinking: are you fuckin serious?

I’d spent the past month rearranging my entire writing workflow. Going back to hand-writing every v1 draft of anything I touch.

At first I figured that came from writing a book. Then the universe handed me a stronger reason.

On August 11, Anthropic confirmed what a lot of people feared: every Claude model released after August 2, 2026 carries an invisible watermark in the text it writes, plus signed provenance data attached to the files it creates. You can’t opt out. You can’t see it. And unless you’re willing to spend two hours rewriting, you probably can’t get it out either.

As an AI-first business owner and writer, I couldn’t ignore this.

So let’s get into it:

  • Why watermarks are a thing to begin with, and why now.
  • How AI watermarks actually work, and what you can do about it.
  • Why it matters if you’re in climate tech, especially if you’re an active AI user.
  • What I’m changing in my own business moving forward.

Why did Anthropic add a watermark to Claude’s text now?

The answer is two more letters: EU.

The European Union has always been the strictest region on privacy and AI regulation. Some call that terrible for innovation. I wouldn’t be so quick to agree — more on that later.

On August 2, the EU AI Act’s transparency obligations became enforceable (Article 50): people have to be told when they’re using an AI system, and when content is AI-generated. That means every AI tool, ChatGPT, Claude, Perplexity, Gemini, has to tag what it produces.

The law is EU-specific. So why is everyone else panicking?

Because separating EU usage from everywhere else means collecting more explicit geographical data, and users would revolt at that. They’d lose customers, trust, and eventually, market share. So the big American tech corporations did the only “sensible” thing: apply the new rules worldwide.

This only takes effect with a new model release, and Anthropic hasn’t shipped one in over two months, so whatever’s next will make a lot of noise online.

To be frank, the internet is already revolting.

But is cancelling your Claude subscription actually the answer?

I’m not a fan of witch hunts, so I dug in, and wasn’t surprised: Anthropic isn’t the first AI company to introduce watermarks. Google has been watermarking Gemini’s text with SynthID since 2024. OpenAI updated its own support page to cover text watermarking in the same week.

Anthropic, OpenAI, Google, Meta, Microsoft and Mistral have all signed the EU code that requires it. Anthropic simply ended up being the beating bag of the bunch.

So no, I wouldn’t cancel your Claude subscription yet. Not until you understand how watermarks work in this specific case.

How does Claude’s watermark actually work?

There are two completely separate watermarks, and only one of them you can actually get rid of.

The first lives in the words themselves, and it doesn’t come off. Claude doesn’t stick a special character on the end or bury something in the metadata; that’d be too easy to strip. Instead it changes the randomness in how it picks each word. Every sentence has a handful of word choices that’d work equally well, and the watermark quietly steers those choices into a pattern its own systems can spot. You can’t see it, and given how much of my own process runs through Claude, I can’t always tell either.

Copy it, paste it anywhere. The pattern lives in the words, not the formatting, so none of that touches it.

But it needs volume to show up. A one-line LinkedIn comment won’t carry the signal the way a whole newsletter issue will. Code’s probably weaker too, since a function only has so many ways to be written before it stops working, leaving less room for a pattern.

If you want the actual mechanics of how that pattern gets built, James Padolsey wrote an interactive essay on it that does the job better than I could in one section.

The second watermark lives in files, and this one comes off easy. Images get signed C2PA metadata confirming Claude touched them. Screenshot it, re-export it, copy it into a new doc, and that metadata’s gone immediately.

Right now that only covers images. PDFs and Word docs are still clean. But it’s baked in at the model level, so switching where you use Claude doesn’t help: chat, Cowork, Code, the API, all of it.

And if you try to edit your way out of the text one? Anthropic’s own words: it “may persist through some editing.”

Why should climate tech founders care about AI watermarks?

Climate tech companies have been fighting greenwashing for decades. It’s cost them time, money, and trust. Three things you don’t get back once a buyer, a journalist, or a regulator decides your claims don’t add up.

So when I say AI watermarks are just another layer of the same transparency push, I mean that literally: this isn’t the first time the EU tried to force transparency behind what a company says out loud.

Back in March 2023, the European Commission proposed the Green Claims Directive. The idea was simple: before you could call anything “sustainable,” “eco-friendly,” or “carbon neutral,” you needed a Life Cycle Assessment and sign-off from an independent verifier. Prove it before you’re allowed to say it.

It never made it. On June 20, 2025, the Commission pulled the proposal, days before final negotiations were supposed to close, after the EPP pushed back on how much paperwork it’d dump on smaller companies.

But withdrawn doesn’t mean gone. Within months, a new version of it was proposed: the Empowering Consumers for the Green Transition Directive, live from September 27, 2026. Generic claims like “eco-friendly” or “carbon neutral” are banned unless you can back them up. Claims based purely on offsetting are banned outright. Get it wrong, and you’re looking at fines up to 4% of your annual turnover, plus a year locked out of public procurement.

Same idea as the withdrawn law, different mechanism. One only wanted proof before you spoke. This one makes you pay for it after, if you didn’t have it.

And AI watermarking is simply another flavour of the same problem: lack of transparency.

Has OpenAI watermarked ChatGPT’s text too?

Not yet, despite what its own announcement implied.

Read the documentation carefully and all OpenAI committed to on August 2 was a future goal: to “expand provenance signals to text models.” A commitment, and we’re all familiar with commitment language in climate. It doesn’t always mean something will actually change.

Which means ChatGPT still doesn’t watermark a single word it writes. OpenAI shipped watermarks for images in May and audio in July.

In fact, OpenAI built a working text watermark back in 2024 and tested it internally at 99.9% effective. But they shelved it because their own internal survey found 30% of users said they’d use ChatGPT less if it shipped a watermark while competitors didn’t.

Whether Anthropic enforcing it now means it’ll fail the same way as OpenAI, or that OpenAI will have to fall in line as well, I’m not quite sure yet. But there’s a precedent, and you can see how everyone’s reacting to the news.

What are the biggest mistakes people make about Claude’s watermark?

Mistake #1: thinking this is an argument against using AI. I use AI every day, in nearly everything I write. I just say so. So do you, and a lot of other founders and professionals in climate tech.

Mistake #2: thinking you can just strip the watermark out yourself. Anthropic already answered that one. Light editing won’t remove it. Only a full rewrite, every word replaced, will — and at that point, rewriting every word is writing it yourself, which defeats the reason to use AI in the first place.

Which brings me to a more practical part.

What does Claude’s watermark mean for your business model?

Tayla Burrell wrote about this same shift and described three potential paths:

1. The “good enough” model. AI-run businesses producing work faster and cheaper than we’ve ever seen: $100 vibe-coded websites, same-day brand kits, accounting at a tenth of what an accountant charges. Nobody expects perfection here, you get what you pay for, but it’s still far better than what most people could make alone. I think an enormous market of micro-businesses opens up in this lane.

2. The “human-made” model. These businesses charge a premium for the one thing getting rarer by the day, that no AI touched it: hand-illustrated brand identity, editorial writing with no AI input, original photography instead of generated stock. You only survive here if you’re exceptional, because you’re competing with people who have far more intelligence and labour on demand than you do.

3. The “AI-first” business model. You take the thing only you have, your IP: the frameworks, skills, methods and insights you’ve built through actually doing the work, and combine it with AI to deliver better results, faster, with less effort. Imagine an AI writing model trained entirely on years of your own writing and principles. Nicolas Cole does it well within the Ghostbase tool.

I’m not a 100% AI-free, human-made writer.

For the past 6 months I’ve been, and plan to continue being, an AI-first business owner: using AI to build more efficiency around my own IP, my writing, my methodology, my philosophy.

In practice: a lot of my research runs through AI, I’d be an idiot not to, it saves a ton of time. I run automations for mundane tasks I hate, like lead scraping and a rough content-analytics run. When I’m tight on time, I dump a large chunk of voice-to-text, the raw context, the idea, the structure, into AI, use my refined skills to generate a v1, then sit down and edit it relentlessly when I find the time.

I also have a 9-11am block every morning now, practicing writing again by hand. I need to keep sharpening the skill: a) watermarking is fairly new and there’s no “proven” way to avoid it, and b) I’d rather get better at writing than try to cheat the system. Forgetting how to write well independently of tools is not an option for me.

If you’re struggling to figure out which direction to go in, ask yourself these three questions.

  • Would you rather focus on a large volume of content and accept sacrificing quality? If yes, you’re closer to good-enough. Nothing wrong with that path, but do expect pushback and a turnoff from readers allergic to AI-generated content.
  • Would you rather every word be unmistakably yours, even if it costs you hours you don’t have? That’s handmade-premium. Expect it to take three, four times as long as anything AI-assisted.
  • Do you want AI’s speed, but retain your authenticity and quality? That’s AI-first, Tayla’s third option, which is where my business is.

The way I see it, the biggest business risk moving forward would be not knowing where you are on that spectrum, and failing to communicate it transparently.

—

I write one of these every Thursday: a short strategic essay on going from invisible to investable as a climate tech founder. The next one lands a week before it shows up anywhere else. Read it first, free: Climate Visibility Lab on Substack.