Understanding AI

The watermark in everything Claude writes

It has been there since the 2nd of August, and right now only one company can read it.

What you'll walk away with

What the watermark actually is, and what it can and cannot reveal about your work
Why it exists, and who the rules genuinely apply to
A prompt that helps you write your own honest position on using AI
The straight read on the removal tools flooding the web this week

About ten minutes to read, twenty if you write your position too.

Work Out Your Own Position

The useful thing to do this week is not to chase a way round the watermark, it is to decide what you actually think. If someone could reliably tell that AI touched a piece of your work, would that be a problem, and how would you answer if they asked you straight out. Paste this into Claude or ChatGPT and it interviews you one question at a time, then hands back a short position in your own words that you can keep.

Paste this into your AI

I want you to help me write my own AI disclosure position. It is a short, honest statement of how I use AI in my work, so that if anyone ever asks me directly whether I used AI for something, I already know exactly what I think and what I would say. Interview me one question at a time, in plain English. Don't write anything until you have asked enough to actually know my answer. Wait for each reply before moving on. Start by asking where AI already shows up in my work, and how much of it. Then ask what I would always be happy to tell a client, an employer or a colleague about that, without being asked. Then ask the harder one, what I would rather not have to explain, and why that is. Then ask what part of the finished work I genuinely consider my own, the judgement, the context, the decisions, the bits that only exist because I directed them. Then ask how I would honestly answer if someone came to me and said, straight out, you used AI for this. When you have enough, write me a short position I can keep. Six sentences at most, in my own plain words, no corporate language. It should say where I use AI, what I am always open about, what I consider mine, and how I would answer if asked directly. Then finish with one thing. Tell me the one or two answers where I sounded least sure of myself, because those are the questions someone else is most likely to put to me, and the ones worth deciding properly before they do.

What It Actually Is

Since the 2nd of August, text Claude writes carries an invisible watermark. Nothing is hidden inside the words. There are no secret characters, no strange spaces, nothing you could find by looking closely. The signal is in which words get chosen.

It is a variant of SynthID, a watermarking technique Google DeepMind published for AI-generated text. When the model is writing and ten different words would all do the same job just as well, it leans, very slightly, toward one of them. One sentence tells you nothing. But stretch that across a whole document and the pattern of those tiny choices forms a fingerprint. It costs nothing, it does not slow anything down, and it does not make the writing any worse. It travels with the text when you copy and paste it into an email or a document, and it survives light editing.

What It Can And Cannot Tell

Reading the watermark needs a secret key, and only Anthropic holds it. So right now nobody else can check. Not your boss, not your clients, not the AI detector you might be tempted to test it against. Anthropic has announced a checking tool that a university or a workplace could plug in, but it has no release date. It is coming, it is not here.

Two more limits worth knowing. It returns a score, a level of confidence, not a clean yes or no. And it needs a decent run of text to work, so a single sentence, or something you have rewritten heavily, will not reliably set it off.

The most important limit is the one people miss. All it can ever say is that Claude wrote this text. It cannot see the hours you spent giving it context, the systems you built so the output actually sounds like you, or the judgement you used to decide what stayed and what went. It flags the tool, not the thinking. Hold onto that, because it is the whole tension underneath this.

Why This Is Happening

This is not Anthropic deciding it wants to police you. It is the EU AI Act, and specifically the transparency rules in Article 50, which say AI-generated content should be marked as such. Article 50 held its 2nd of August date. The marking obligation for AI models that were already on the market before that date was pushed back to the 2nd of December 2026, so Anthropic did not have to move yet. They went early.

Which means the rest are not far behind. Every major provider faces the same rule within months, so within the next while, this quietly becomes the normal state of using any AI to write.

Can You Remove It

This is the question everyone lands on, so here is the honest answer rather than a tool, and a quick note that this is my plain-English read and not legal advice. As the AI Act is written, it aims its obligations mainly at the companies providing and professionally deploying these models, rather than at a private individual writing something at home. So for most people using AI for their own personal work, this particular law is not the thing to worry about.

For a business publishing content, it can be a different picture, and this is where it is worth getting proper advice rather than taking mine. Deliberately stripping out a transparency mark is the kind of thing a transparency regime is designed to discourage, and the potential penalties are not small. That is a real risk to weigh, not a technical puzzle to solve.

And the watermark remover tools that have appeared across the web this week share one problem. Because only Anthropic can currently read the mark, nobody selling a remover can independently prove theirs works. So anyone telling you to comment this or download that to make your text undetectable cannot really back the claim up. I am not going to hand you a method, because there isn't an honest one to hand.

The Honest Bit

I genuinely do not know how I feel about this one, and I am not going to pretend otherwise. Part of me bristles. I have spent months building systems so that what comes out is a fair reflection of what I actually think, and the only reason those words exist at all is that I prompted them into being. The other part of me thinks being straight about where AI helped is just reasonable, and that hiding it is the thing that would sit wrong.

I am not going to tell you where to land on that, because it is genuinely yours to decide. The useful move is the one at the top of this page. Work out your position before someone asks you for it, so that if the day comes and a client or an employer says you used AI for this, you already know exactly what you think and how you would answer.

One bound on all of this. I run content, not a law firm, so none of the above is legal advice. If your business has real exposure here, the enforcement sits with national market surveillance authorities and the numbers are large enough that it is worth a proper conversation with someone qualified.

Where This Goes Next

The whole ownership point rests on one idea, that the work is yours because of the context and direction you put into it, not the tool that typed it up. The context is what your agent actually knows is where that starts, and why your chat gets worse the longer it runs is the mechanism underneath it. For the rest of these write-ups, head back to the resources.