AI content is getting a paper trail. That's only a problem if nobody owns it.

Yesterday I pasted a paragraph I'd written myself into Claude and asked it to tighten the phrasing. Nothing unusual. I do it most days.

What came back was apparently watermarked.

Not visibly. There's a pattern woven into the wording that a detection tool can pick up, and it travels with the text when you copy and paste it. My words, my argument, my thinking, and a mark on the way out.

I'm not an AI expert and I'm not going to write like one. I use these tools daily, inside a fairly tight set of rules, and I've spent a week reading about this properly because it changes something about how AI-assisted work behaves once it leaves your building.

Short version: if you work with your own judgement and a real review step, this is a non-event. If AI has been your volume machine, you now leave a trace.

What's actually happening

Anthropic confirmed it on 11 August. Claude now embeds an invisible mark in the text it generates. Anthropic says it doesn't change the meaning or readability, it survives copy and paste, and it may survive some editing. Nobody will say how much editing removes it.

Files work differently. For supported types like .png, .jpg and .svg, Claude attaches signed metadata using C2PA, the open standard the industry is settling on. That shows the file was processed and whether it's been altered since. It's also easy to lose: a screenshot or a format conversion strips it.

It isn't retrospective and it isn't everywhere yet. Models launched from 2 August 2026 carry the mark from launch, older ones are being retrofitted, and systems already on the market have until 2 December. Not every platform supports every type of mark.

And this isn't one company's decision. Google, OpenAI, Meta, Microsoft and Mistral have all signed the same EU Code of Practice on Transparency of AI-Generated Content, though each can implement it their own way. It sits under Article 50 of the EU AI Act, which applies from 2 August 2026 and asks providers to make AI output detectable.

It isn't only about AI-written content

Everyone read this as “AI writing gets tagged”. That's not quite it.

A mark can appear on work the model only handled. Proofreading. Translation. Summarising. Converting a file. Which is what happened to my paragraph, and it's what most firms actually use these tools for.

So a board paper you wrote and tidied up can carry the same basic signal as something spun out of a one-line prompt: that Claude may have touched it. Same for the case study your engineer drafted and someone polished on a Friday afternoon.

The thing is that the law didn't ask for this. Article 50 carves out systems doing routine editing, where the tool isn't really changing what you said. Tidying up your own prose looks like exactly that. Anthropic seems to have gone further than it had to.

Which tells you something useful about the next couple of years. If you set your expectations by what the law requires, you'll be wrong about what the tools actually do. The vendors are already ahead of it, and they'll probably keep going, because being able to prove where content came from is turning into a feature rather than a cost.

A mark proves less than you'd think. So does its absence.

Anthropic's own documentation is unusually straight about the limits, and worth reading rather than taking from headlines.

A detected mark means the content may have been processed by Claude. It doesn't tell you who wrote it, whether the thinking was original, or whether the text changed afterwards.

Absence proves nothing either. Heavily edited, paraphrased, translated, mixed into other writing, or just short, and there may be nothing left to find. Older models aren't marked yet.

Detection tools haven't shipped. Anthropic says they're coming and hasn't put a date on it, so nobody is running this over your website this afternoon. I wouldn't guess at when they will be.

This puts the mark somewhere useful and unglamorous. It's a signal, not a verdict. Anyone treating it as proof of laziness will be wrong a lot. But it exists now, and it didn't in July.

Most of this isn't about you. The narrow bit where it is.

Article 50 does ask firms using these tools to label AI-generated text. But only where three things are true at once. It's published. It's meant to inform the public. And it's about a matter of public interest.

The Commission gives examples of what that covers: politics, public administration, justice, fundamental rights, public security, public health, environmental protection, consumer safety, and financial, scientific or cultural developments people might reasonably debate. Examples, not a closed list.

Notice what isn't in the test. Format. It makes no difference whether the thing is called a case study, a product page or a blog. A case study about a CRM migration is nowhere near any of those categories, and neither is most day-to-day B2B marketing. Anyone telling you to stamp an AI label on everything you publish has over-read this.

But a case study about building safety, energy performance, product compliance, environmental impact or workforce health is a different proposition, and it looks identical on the page. Firms in regulated sectors build authority by publishing on exactly those subjects. That isn't a fringe activity. It's what credible thought leadership looks like across half the B2B world, and it's where the boundary stops being obvious.

Then there's the exemption, which is the part worth reading twice. Text that's been through proper human review doesn't need a label, as long as someone carries legal responsibility for publishing it. The Commission is specific about what review means: somebody with real knowledge of the subject looking at the substance. Spell-checking and grammar fixes don't count.

So the test isn't whether AI touched the work. It's whether somebody who knows the subject read it properly and put their name to it.

I wrote in July that AI is good at scaffolding and bad at judgement, and that the firms getting value from it know which is which. I didn't expect the law to land in roughly the same place a fortnight later.

One disclosure, since it would be odd to write this piece without one. I used AI to get through Article 50 and the guidance around it, because they're long and I'm not a lawyer. Then I checked it against the Commission's own pages and found I'd had the scope wrong. The primary source corrected me, not the summary. That's the whole argument in miniature, so it seemed worth saying rather than quietly fixing. Sources are at the bottom. Anything still wrong is mine.

The question was never who typed it

There's a moral temperature to this conversation that the facts don't support.

Professional work has always separated who produced the words from who is responsible for them. Chief executives give speeches they didn't write. Boards issue statements drafted by advisers. Partners sign reports assembled by teams. Nobody calls it fraud, because a name has never claimed to identify the typist. It identifies who owns the position.

That distinction is old, and it's the one Article 50 reached for. Not “was a human at the keyboard” but “did someone competent take responsibility”.

Which is why the objection to AI has been aimed slightly off-target.

There are places we do care about authorship. Sworn testimony. Academic work. Investor communications. But look at what's being protected. Not the typing. It's that the person whose name is on it understood it, meant it, and can be held to it. Substance and responsibility, again.

And at the other end sit all the tools nobody has thought about in years. Nobody asks whether you used spellcheck. Nobody asks the accountant whether they reached for a calculator or did it longhand. Nobody asks whether you knew the route or let the sat nav handle it. Those arguments, if they ever really happened, are long over.

What they have in common is that it became obvious the person was still answerable for the result. A sat nav is no defence for turning up at the wrong office.

We've also used effort as a rough proxy for care. If someone spent three hours on it, they took you seriously. AI weakens that, because the output no longer tells you how much attention went into it. That's a real loss and it's why the unease isn't silly. But the thing being protected was never the effort. It was knowing a competent person had actually thought about your problem.

That's still available. It just has to be given directly now, rather than assumed from the fact that something looked laborious.

Where the reassurance runs out, and I'd rather say it than have you think it, is that spellcheck couldn't be confidently wrong about a load path or a compliance clause. It caught the typo or it didn't. AI produces things that look like judgement and occasionally aren't. Which is why the review step isn't optional, and why “it's just the new calculator” is a bit too comfortable.

So this isn't a moral question. It's a quality control one. They need different answers, and firms keep reaching for the first when they need the second.

So what do you actually do

Not much, if the house is in order. Three things worth checking. ‍

Someone's name is on it. Legally, responsibility can sit with the company. Practically, “the company reviewed it” means nobody did. Every piece of external content wants a person who read it and would defend it in a room.

The review is real. Someone who knows the subject looking at the substance. Not reading for typos. Someone who can tell whether the claim is true, the emphasis is right, and it sounds like your firm rather than any firm.

That can be the same person who prompted it. But they have to come back as a critic rather than as the author, and that's harder than it sounds. You are reliably the worst reader of something you just watched appear. Leaving it overnight does more than most review processes.

Your AI policy covers the way out, not just the way in. Most policies are about what staff can paste in. Far fewer say anything about what happens to the output. Half a day's work, and it's now the half that matters.

Two places I'd pay closer attention

If you sell software with AI features built in, you may have obligations of your own rather than just the ones that come with using the tools. A firm putting an AI assistant in front of customers is in a very different position from one using AI to write its blog.

Being outside the EU doesn't settle it either. The Commission is clear that firms based elsewhere can be caught where the output is used in the EU. That's a legal question rather than a marketing one, and it belongs in front of someone qualified rather than worked out from articles like this one.

The second one is quieter. If you sell expertise at expertise prices - consultancy, engineering, professional services, agency work - your exposure isn't regulatory. It's commercial. The awkward conversation isn't with a regulator. It's a client running a detection tool over a deliverable they paid senior rates for.

That's fine if you've already been open about how you work. It is not fine as a surprise. Which is a decent argument for saying it out loud now, while saying it is still a choice.

What I'm doing about it

Nothing different, which is the honest answer.

I use AI daily on the scaffolding. First drafts. Research I then check against primary sources, as this piece rather proves. Turning an hour of messy notes into something legible. The judgement stays mine: the positioning calls, the read on what a particular buyer will and won't believe, the bit you're actually paying for. My name goes on it either way.

My position hasn't moved. Use it inside a structure, with your own expertise doing the deciding and a named person carrying the output. Then a mark on that work tells the world something you'd have been happy to say anyway.

The firms who should be nervous are the ones who wouldn't.

‍If you're deploying AI across your marketing and you'd rather build the review structure now than retrofit it later, that's a conversation worth having. clare@clarewatson.uk

‍ ‍

Sources: Transparency obligations under Article 50, European Commission  ·  Article 50 of the AI Act, full text  ·  How Claude marks AI-generated content, Anthropic

‍ ‍