There was a lot of noise this week about Claude watermarking its own text. And the panic was all one note.

“Oh no. People are going to find out I used it.”

That reaction tells you everything. Nobody has ever panicked about being caught using a spell checker.

What the watermark actually proves

Anthropic now weaves an invisible mark into text Claude generates. It nudges word choices into a pattern you can detect over enough text, and it survives a copy and paste. It applies everywhere, and it’s happening because of the EU AI Act rather than because anyone had a moral awakening.

All of that got covered to death this week. What didn’t was Anthropic’s own help page on it. They say a detected mark means the content may have been processed by Claude. Then they say this outright: “Claude may not be the original author. People often use Claude to proofread, translate, summarize, or convert files. The output can carry a Claude mark even if the underlying ideas, text, or data originated from another source.”

The watermark cannot tell you who did the thinking. It tells you the text went through a machine at some point. That’s it.

So everyone panicking is panicking about the wrong thing. The question “did you use AI?” is already dead, and not because of watermarking. It’s dead because loads of companies have bought these tools and are actively telling their staff to use them. That argument is over. You’re not going to get in trouble for using the thing you were told to use.

The question is whether you’re using it well.

Or more bluntly: what did you bring?

If you feel guilty, that’s useful information

I do think some of that panic is real, though. And I don’t think it’s about detection.

If you feel a bit sick about someone finding out, it’s usually because the model did the thinking you should have done. That guilt is data. Listen to it.

Because there’s a version of this where you have nothing to feel bad about at all.

Take this blog. I record myself babbling into a voice recorder on my Mac for a few minutes. That’s where it starts. The thinking, the angle, why I even want to write about it, all of that is mine. I do the research and I get it double checked. Not to mention the hours of just general thinking I have been doing in my head over the past week.

What the model does is take that mess and write it up. I don’t think I’m a good long form writer. That’s exactly where it helps me, and I’m fine saying so.

Work is the same. I don’t open Microsoft Copilot and say “make me a presentation about X”. I plan in text files first. I get the argument straight, go back and forth to build a more robust draft, and only then does a first-draft PowerPoint fall out of the other end.

The problem is that nobody is going to trawl through my text files to work out what I added. The proof exists. It’s just completely unreadable.

Nothing we send anyone says who did what

This is the actual gap, and it isn’t a detection gap. It’s a provenance gap.

Notice we’ve now got two separate systems for tracking the machine. The watermark handles text. Images get C2PA, and that one isn’t an Anthropic invention at all, it’s an industry standard with thousands of organisations signed up to it.

Adobe writes it across Photoshop, Lightroom and Firefly. Leica, Sony, Nikon and Canon sign photos in the camera at the moment you press the shutter. OpenAI stamps it on its image output. Google and Meta both read it to put “AI info” labels on what you’re looking at. Anthropic attaches it to generated image files too.

It’s a serious piece of work and it does its job well. It records that a machine touched this file, and whether anyone’s altered it since.

So: two mechanisms, years of industry effort, thousands of companies. All answering the same question. What did the machine do?

There is nothing that does the same for the human.

I kept circling round to git. In code you can see exactly what changed, when, and by whom. Beautiful. But nobody is going to read a commit history to decide whether your deck was worth what they paid for it. A full version history is proof that nobody will ever look at.

So the useful version has to be small.

People have tried this, and they’ve all built the same thing

I should say up front that I’m not the first person to notice this. Go looking and you’ll find a few attempts.

The most developed one is AIUC, the AI Usage Classification. Five levels: AI-Free, Human-Led, Co-Created, AI-Led, AI-Generated. You get issued a badge, and they compare it to a nutrition label, which is a decent way of thinking about it. There’s a code of practice behind it. There’s a Human Standard Institute doing something similar.

And it’s not just a nice-to-have any more. In the US there was an FTC policy statement in March that put companies at risk of “unfair or deceptive acts” charges for claiming human-made quality on AI-produced work. So this is coming from the regulators as well.

But they all do the same thing. They label the whole artefact on one scale.

One badge for the entire deck. And that collapses the exact thing I care about.

Run this blog post through it and you’d probably get “Co-Created”. Technically true, but completely useless, because it hides the only interesting fact about how it was made: the direction is entirely mine and the drafting is entirely the machine’s. A single badge can’t tell you that. It averages me out and hands back a number.

Which is precisely the problem academia solved in 2014.

Put a HAT on it

So here’s my crack at a different shape. I am calling it HAT or a Human Attribution Tag.

A few lines at the end of a piece of work saying who did what. A last slide, an appendix, a block at the bottom of a post. Not a legal disclaimer, not a hedge, not an apology. Just a statement of what you brought.

And I want to be clear that the idea is borrowed, not clever.

There’s a thing called CRediT, the Contributor Roles Taxonomy. It’s a properly ratified standard, ANSI/NISO Z39.104-2022, used by more than 50 publishers across thousands of journals. Research had the same problem we do: one author credit on the front of a paper told you nothing about who actually did what. So they binned the binary and replaced it with named roles. A paper now tells you who did the conceptualisation, who did the methodology, who did the writing, and who did the review.

Not a badge. A breakdown.

It’s been running for over a decade, nobody thinks a paper is weaker for carrying one, and it never left research.

So let’s learn from it. CRediT has 14 roles, which is far too many for a Tuesday afternoon deck. I’d start with five, and each one gets marked as human, AI-assisted, or AI-generated:

• Direction: the argument, the angle, what this is even for
• Research: gathering the facts and checking them
• Drafting: turning the plan into prose or slides
• Review: actually reading it properly
• Sign-off: who is accountable when it’s wrong

On a real client deck it might look like this:

HAT - Human Attribution Tag
Direction:  Human. Planning file dated 12 Aug, pre-dates any AI involvement.
Research:   Human-gathered, AI-assisted summarising. Sources in appendix.
Drafting:   AI-generated from the human plan. Images AI-generated.
Review:     Human. 3 passes.
Sign-off:   [name]

That took about a minute to write.

What would make this stick

If this is ever going to be more than a blog post, I think it needs five things.

• A name you can say out loud. Half of why CRediT worked. “Put a HAT on it” is short enough to survive a meeting.
• Few roles, and fixed. At 14 roles it becomes a form, and forms get ignored. Five is probably the ceiling.
• A predictable place to look. Last slide, document footer, end of post. Where it lives matters more than how it’s formatted. The value is knowing it’ll be there.
• Plain text first. Machine-readable can come later. Anyone can adopt this tomorrow with no tooling and no budget.
• Someone with leverage has to ask for it. Journals didn’t adopt CRediT voluntarily, publishers required it. At work that’s whoever owns the deliverable template. One slide, not a procurement project.

I don’t know if this is right. It’s a first crack and I’d like people to poke holes in it. I know I am going to try and see how it goes.

The obvious problem

Someone is going to say: anyone can just lie in the tag.

Yes. Obviously. People lie on their CVs too and we still have CVs.

It’s still worth writing, because being asked to state it plainly makes you notice when the answer is embarrassing. If you have to type “Direction: AI-generated” on a piece of client work, you feel that. That’s most of the value right there, and it happens before anyone else ever reads it.

I did think about publishing the raw transcript of this one as proof. I’ve decided not to. It’s a voice memo of me babbling for seven minutes. Letting people that far into how messy the inside of my head is feels like a step too far, and the tag does the job without it.

I got the machine to write mine, and then argued with it

I wrote a HAT for this post early on, from memory, while I was still drafting. Direction: Human. Research: AI-assisted. Felt about right.

Then it occurred to me that I was the worst possible person to be filling this in. I’m marking my own homework, from memory, about my own contribution. Nobody is honest under those conditions, not deliberately, you just round yourself up a bit without noticing.

So I built a prompt to do it instead. Feed it the work and what actually happened, and it produces the tag. Partly because it’s less flattering than I am. Mostly because if this is going to be a thing people do at the end of every job, it has to take thirty seconds, and me sitting there weighing up my own worth is not a thirty second activity.

Then it disagreed with me, which I wasn’t expecting.

It was right about Research. Not AI-assisted, AI-led. I didn’t find the line in Anthropic’s documentation this whole post hangs on, and I didn’t know CRediT existed. What I did was push back twice, once on whether C2PA was the right reference and once on whether anyone had tried this before, and both pushes changed the piece. Real contribution. Just not the one I’d written down.

Then it came for Direction, and that one I fought.

Its argument was that “borrow CRediT” and “badge versus breakdown” weren’t my ideas, so the direction was shared at best. Except I went back to the voice memo. And there I am, before any research happened, saying we’ll need to “add something to the end, like a last slide or an appendix, what did an LLM do, what did the human do.” A minute earlier I’m comparing it to git, where you can see what changed and by whom.

That’s the idea. Badly said, but said.

What the research did was hand me a name, a ratified standard, and proof that a decade of academics had already made it work. That’s not the idea. That’s evidence for the idea. The difference matters, because if it doesn’t, then anyone who looks something up owns your thinking.

So the honest answer wasn’t the flattering version or the harsh version. Finding it took an argument and a trip back to the source recording, and I only won because that recording exists.

Receipts.

So, what did you do?

The watermark is coming whether any of us like it or not, and it proves a lot less than the headlines suggested.

I’m not remotely worried about someone reading this and going “an AI wrote this.” Yeah. Partly. And now you can see exactly which parts.

We’ll get bored of asking whether an LLM was involved, exactly the way we got bored of asking whether someone ran a spell check. What’s going to stick around is the question underneath it.

What did you do?

HAT - Human Attribution Tag
Direction:  Human. 7 minute unscripted voice note set the argument, the
            per-role idea and the conclusion, before any research.
Research:   AI-led. Supplied the CRediT precedent, the AIUC prior art and
            the "badge vs breakdown" framing. Human challenged it twice,
            correcting a C2PA error and surfacing the prior art.
Drafting:   AI-generated from the voice note.
Review:     Human. 5 passes, substantive cuts and factual corrections.
Sign-off:   Naji El-Arifi