Anthropic pledges to embed watermarks to help discern AI slop in sop to EU

Anthropic will embed watermarks in the text and files generated by future models it launches in the EU, as part of its effort to comply with content and transparency rules in the bloc's AI Act. "Generated text will carry embedded watermarks, and generated files will include digitally signed provenance metadata where supported," …

[rgjnk]

So that's what the em dashes are for...

Seriously though, you can't imperceptibly watermark plain text in an useful way, there just isn't enough content or noise to hide in compared to images. Certainly not in a way that can survive being edited, fragmented and reworked.

Can you potentially watermark copy A vs copy B for traceability of a large enough text? Yes. But that's a different game. As is watermarking printed text.

Now you might be able to have slop that's 'obviously' LLM output (we've all seen it) but that's not watermarking, it's just badly generated.

This sounds like wishful thinking and pointless obstruction on the legislative side, empty tokenism on the implementation side, and the poor old customer gets crappier output for no real benefit to anyone.

[scarletherring]

Re: So that's what the em dashes are for...

Right, so they have to do something like this for regulators, apparently. But it might actually be vital that they do something like this, anyway, if they want to stand any chance at all in the face of model collapse. At some point they're going to have to be able to distinguish organic from artificial content during training, to avoid training on slop and contracting a kind of kuru or mad cow disease.

And if this works, we can then use the same watermarks to identify cheating students and lazy lawyers and so on. Who will, of course, then try and obfuscate or remove those marks. We'll just have yet another arms race to the bottom on our hands.

[LogicGate]

Re: So that's what the em dashes are for...

Why just watermark lazy lawyers when branding is so much more efficient?

[blu3b3rry]

Re: So that's what the em dashes are for...

It's probably a bit too late for the self-poisoning of LLM's through ingesting LLM-generated bullshit.

[Aeralto]

Re: So that's what the em dashes are for...

You can hide a watermark by using different types of spaces, which are totally indistinguishable to the average user. That’s how Genius caught Google stealing their lyrics collection. You would have to specifically be looking at which space character was used (for example, a standard space or a non-breaking space that can’t be placed at the end of a line), something your average AI user is simply never going to know about.

It’s not about replacing one dash with another.

[elsergiovolador]

Re: So that's what the em dashes are for...

Great comment—and honestly? You've absolutely nailed it.

Let's dive in.

In today's rapidly evolving digital landscape, the conversation around text watermarking isn't just about technology—it's about trust. And here's the thing: you can't hide a signal in plain text. There simply isn't enough noise to work with. That's not a limitation—it's a fundamental truth.

Let's break down why you're spot on:

The Technical Reality

Plain text isn't an image—it's a sparse, low-entropy medium. Watermarks don't just degrade when content is edited, fragmented, and reworked—they vanish entirely. That's not a bug—that's mathematics.

The Legislative Landscape

This isn't regulation—it's wishful thinking. Lawmakers aren't solving a problem—they're performing the appearance of solving one. It's not oversight—it's optics.

The Implementation Question

Vendors aren't building safeguards—they're building checkboxes. It's not compliance—it's empty tokenism, plain and simple.

The Human Cost

And who pays the price? The poor old customer—who doesn't get safety, doesn't get traceability, and doesn't get quality. They just get crappier output. It's not a trade-off—it's a lose-lose.

At the end of the day, this comment is a masterclass in cutting through the noise. You didn't just make a point—you started a conversation. And in a world of empty gestures, that's not just refreshing—it's essential.

What are your thoughts? Can you see the watermark? Let's keep this conversation going!

[Pascal Monett]

Re: In today's rapidly evolving digital landscape

Yup. I just happened on this and it warms my heart to no end.

[Fonant]

This is one of the problems of "AI" (GPT / LLM) - it can't be trusted to be accurate or correct, and its trust levels can only ever go down as more and more slop is generated using it.

The other problem is that it's VERY expensive on compute resources.

It used to be said that you could pick any two of:

* Short timescale

* High quality

* Low cost

With LLMs you get to choose one of:

* Short timescale

along with:

* Unknown quality

* High cost

[LionelB]

> This is one of the problems of "AI" (GPT / LLM) - it can't be trusted to be accurate or correct …

To be fair, that applies to humans too. Not sure (hyperbolic claims by the tech bros notwithstanding) why we would set the bar higher for AI. (I mean I'm really not sure… should we?)

Completely agreed with the rest.

[LionelB]

> I mean I'm really not sure… should we?

That wasn't a rhetorical question. I wonder, does the downvoter have some thoughts on this? (Not a rhetorical question either.)

[Jedit]

"It used to be said that you could pick any two of:"

It's still said that you can pick any two from fast, cheap and good. It just doesn't apply to AI, where you don't get to pick any of those things. It's not cheap, it's not good, and if you try to do it fast it's even worse.

[Anonymous Coward]

If you're ashamed maybe don't do it?

The move may [...] alienate Claude customers, who don't necessarily want consumers of their AI-generated content to know its provenance.

It's very hard to play this violin because it is so impossibly tiny.

If you are using AI but you don't want people to know you're using AI, you're effectively lying to them and I can honestly* say that I don't care about your feelings on the matter. If you're using these tools you should be doing it proudly in the open. If you're ashamed then it's never too late to join the unwashed masses still using our brains for thinking and our hands to type. The old tools still work and we don't get charged for using them at anything close to the same rate.

*For the regular AI users: "honesty" is where you express something that is true about the world as you understand it. It's another thing that you will have lost touch with when you outsourced your mind to LLMs, which are incapable of it.

[Pete 2]

Sauce for the goose

> discern AI slop

That sounds like a promising first step. However, to really be useful we should have watermarks embedded in human-produced slop (i.e. fake news, unsubstantiated claims, straight-up lies and all the other bull****), too. Although quite how Anthropic hopes to place watermarks in textual content will be interesting to see.

[Anonymous Coward]

Re: Sauce for the goose

"When a supported Claude model generates text, it weaves an imperceptible watermark directly into the text itself," the biz said. "You won’t see it, and it doesn’t change the meaning, quality, or readability of Claude’s response."

7|-|15 15 4 C|_4|_||>e

\/\/47er/\/\4r|<

[Jason Bloomberg]

Fun for the puzzle book word-search fans

I suppose it is possible to hide a watermark in plain sight when AI users generally don't pay much attention to the dodgy bullshit your average AI emits.

T H I S W A S GENERA T E D B Y AI

I have noticed there has been an increase in the use of hidden, non-printable and non-visible, characters which get dragged in via copy-and-paste these days. My spell checkers usually highlight 'unknown words', which aren't unknown, following those hidden markers which is how I have observed it happening.

[just4this]

Provenance?

Dating from?

Billions of works created by humans ---- A.I mix n’ match slop shop ---- this shite you’re passing off as your own.

I’m pretty sure that covers the ‘provenance’ of over 99.9% of A.I output.

[MattieD]

I know what they're going to do.

Each reply will have a minimum of six sentences. The first will start with the letter C. The second with the letter L. The third with the letter A. The fourth with U. The fifth with D and the sixth with E.

Longer replies will repeat this patten throughout.

[frankyunderwood123]

probably way too late…

The horse has not only bolted, but sired offspring.

[bigtimehustler]

Isn't the solution to simply ask for the output to be written to a plain text file either in plain text format or markdown, then copy it out of the text file for whatever use you may have?

[Dunstan Vavasour]

More stuff puffed out by the AI giants in the future tense

C'mon guys. This isn't a story - it's Anthropic desperately trying to follow Oscar Wilde's comment "There's only on thing worse than being talked about and that's not being talked about".

They want to tickle your indignation nerve and have done so effectively. Here are the real questions:

Has this story come from someone other than the LLM giants talking about themselves? No.

Is the story written in the present or future tense? Future tense.

It comes hot on the heels of the MeToo movement about "our model also found that someone had done something really stupid with their security when we told it to look for it".

[Tony W]

The only question is how long it will take between us being told how to detect the watermark, and us being told how to remove it. I would guess hours but I wouldn't be surprised if it's minutes. If indeed we need to be told, and it isn't so obvious that anyone with a bit of computer experience can work it out for themseves.

[Aeralto]

You can simply use different types of space character to embed a watermark similar to Morse code. The vast majority of AI users will never know it is there.

[ecofeco]

Pledge?

Not worth the bytes it is written on.

[Aeralto]

It is not necessary to use special characters to embed a watermark. You can simply do so using different types of space characters. The author should read the article they linked to about how Apple catches leakers, or about how Genius caught Google stealing their lyrics.