This should be legally mandatory by use in countries.

As I’ve said in response to a number of articles about this, all this means is that when someone wants to use Claude’s output for something, they just need to remember to manually type it in now, rather than just copy-paste it. Practically speaking, this changes nothing.

It’s some kind of statistical thing about what word is where, what character is there how often. Typing won’t help, but a rewrite will.

What I’m saying is that if someone just types verbatim what Claude says, instead of just going “Ctrl-C, Ctrl-V”, that completely undermines this whole thing.

You seem to be assuming that they’ll be embedding hidden characters or something. That is not the case; the watermarking will be based around West specific words are chosen, meaning that if you type in those exact words, you’ll likely preserve the watermark. There are more details here: explainx.ai/…/anthropic-claude-invisible-watermar…

Very interesting read, thanks for sharing.

I dont think, he’s not hiding any invisible stuff in ASCII, it’s not technically possible. But the pattern can itself be a watermark like the person before me just said

Their point is that Claude is now engineering its responses to specifically conform to a detectable pattern, something that can of course be easily removed via a seperate LLM.

That was the author’s guess. Seems like this would only be detectable if it’s being used in enough text (possibly across multiple files).

The difference between auto-generating hundreds to thousands of words, versus having to manually read and type all of those words out yourself is very obviously a lot more than “nothing”.

And that’s not even taking into account the fact that we don’t know what their method of watermarking is. From how they’re talking it sounds like it will be algorithmically embedded in the text itself somehow, through specific choices of words or something similar. If that’s the case, typing it out word for word will retain the watermark.

How tho? Will Claude be looking over our proverbial shoulders now, and if whatever we type just so happens to be word-for-word what it said, it can magically flag it?

The words and the text itself are meaningless, it’s the content that matters

You have a fundamental misunderstanding of how the watermarking works. There is no hidden metadata, but most likely, the statistical properties of the generated text are being changed.

If you met someone on Halloween with a face mask, and they’d always use the word cromulent in every sentence, you’d probably assume it’s your buddy Mark, who is about the only guy in your circle of friends who does that.

The model will produce a text where individual words at certain positions, or various n-grams encode a kind of fingerprint that will be an indicator for the text being processed by Claude.

Like, when people have, like, a specific accent or talk in a certain way, you can totally, like, figure out where they’re from, for sure.

So isn’t the solution then to just feed the Claude output through a second tiny LLM with the prompt to slightly rewrite the input? I mean you can probably use a 1b model to do that and that can run on practically anything nowadays

Yes, if the watermarking works like we all assume, significantly editing the output will likely break the watermark. They say this in their press release, as well as that the text must be a certain length to be watermarked. (Probably as a function of how the watermark works.)

I don’t think this is expected to be 100% foolproof. They certainly don’t make that claim; they explicitly say that the lack of a watermark isn’t conclusive evidence that the text wasn’t generated by their models.

There’s typically a keyboard shortcut to copy things sans formatting (plain text.)

I don’t think you will even need to type yourself. Tools removing the watermark will be available 3-4 days after this hits the market, fully able to be automated. Watermarks are not a solution that will prevent someone who wants to deceive you from doing so - only awareness that everything on the web is sus will.

Edit: The only alternative is an 100% digitally signature authenticated web, from the web login down to the posting. I am not sure this is a better way because it also means 100% transparency of everything someone does.

Correction: consumer grade machine-extruded text will be watermarked.

Machine-extruded text generation controlled by government entities, fascist regime propaganda outlets, capitalist interests and basically anyone else with a ton of money will be welcome to bulldoze venues of public debate with as much un-watermarked thought-slop as they care to generate that will be virtually indistinguishable from the real thing.

Imagine layers of public influence campaigns, all conjured by LLMs, with agents A/B testing rhetoric constantly for success and efficacy.

Call me cynical, but I think this is an attempt to do the bare minimum in an effort to preempt laws being put in place.

I think they want to detect others distilling their models.

And enshittify their “basic” models so they can charge premiums to not watermark text.

The year is 2027. A beginner programmer searches how to access the last array element in a programming language she is unfamiliar with. The results come in. “Top 10 array accesses that will make you want to buy our sponsors product”. She sighs and asks Claude instead. She pastes the one-liner into her code. Finishes the remaining code. Tests. Double-check. Commit. Push. A notification appears. “Your account has been temporarily suspended for violating GitHub’s terms of service”. Forgot to mark claude as a co-author. She sighs again

Or the ToS violation will be that she didn’t use Copilot (which will probably have a different name in 2027)

Microsoft AI Explorer One

It’s that the one better or worse than the Explorer One S?

ClippAI

I hate how accurate that is

Microsoft Windows X-360+5

Introducing Microsoft Copulate. Our new AI agent that will fuck you over in more ways than one

These watermarks require the text to be long-ish and not too edited from the original. I didn’t see this explained in the article, but I only skimmed it. The press release from Anthropic goes into it, without giving many details.

I’m guessing it’s going to embed a pattern with punctuation and word/letter choices, so there will need to be a sufficient amount of text to ensure its actually the watermark and not just dumb luck.

Yeah it’s exactly a pattern with token choices, by introducing a statistical bias to the randomness in token generation. Quality isn’t affected because the model is still choosing from the best candidates for the next token. You’re right that to be effective they need a long enough input to get enough matches to be.

Computer code is also going to be less effective because code has a much more rigid structure dictated by language syntax rules, coding conventions, linters and formatters, etc.

They want to check and control what the output is used for, IE who train on their output… The big US scum-corps are just tightening their ps*cho control regime (although most of us escapes them via Chinese open models)…

Those hidden markers will be very hard to hide in code review in an IDE which highlights invisible Unicode code points. For code, adding watermarks is likely to cause bugs. For text, Chinese models are as good - so just switching to them solves the issue.
I think, Anthropic doesn’t want to perform economic seppuku.

These are not watermarks using hidden characters. Their approach is undetectable even with an IDE. For a sequence of tokens, an LLM predicts the most likely next token, with some amount of randomness between equally likely candidates. The “watermark” is to introduce a statistical bias to this randomness, by altering the probability distribution of generated text according some hash function with a secret key, thereby embedding a statistical signature into the text itself.

So if the text is “I like to eat __” the model might have 3 top candidates for the next word (apple/orange/banana) that would be chosen at random. Instead that choice will be biased towards one option according to their hash function. And then again “I like to eat banana __” (cake/pie/tart).

To verify a text, they look for the “watermark” by scanning the text and looking at whether sequences of tokens chocies fits their biased probability distribution or are truly random. Just one match doesn’t tell you anything, but if they see a consistent pattern over a 1000 word document, they can give a very high confidence that their model generated the text.

To a human it looks like nornal generated text, and the output quality isn’t affected much (or ar all). It’s much more effective on generated prose, and not very effective on computer code.

I mean it kind of already does this accidentally, just look for em dash and/or emojis sprinkled all over the document and you can be pretty certain that it was AI generated.

The other thing I’m seeing here is that this will only be effective for large chunks of text, if you’re dealing with small snippets interspersed with human generated content there will be enough statistical noise to make classification hard without introducing a bunch of false positives and negatives.

That sounds a bit like it conflicts with the actual job of the LLM. And the resulting watermark would be way too fuzzy to be actually useful for flagging anything as AI-generated.
Are they just trying to check a box on some compliance checklist?

That sounds a bit like it conflicts with the actual job of the LLM.

You’re right. But it’s designed in such a way that it only biases the choice between the statistically most likely candidates, so it’s not forcing a choice to a less optimal token. It’s biasing the choice between equally optimal tokens. So it doesn’t really affect the quality of the LLM’s output.

And the resulting watermark would be way too fuzzy to be actually useful for flagging anything as AI-generated.

It’s actually not that fuzzy. It’s the statistical equivalent to randomly guessing a 256-bit encyrption key. If you consider the algorithm operate on trigrams (sets of 3 words), then a 1000-word document contain 998 trigrams. Let’s say at each trigram the model has a choice between 16 equally likely candidate words, which is usually chosen at random according to the model temperature, now is also biased by the watermark hashing function.

The statistical likelihood of randomly making the same 1/16 choice as the watermark 998 times in a row is so extremely small it’s essentially impossible. Even if you rearrange the document, cut large portions, paste in other portions, rewrite some, you’re likely to leave in enough matching trigrams to make a statistically solid determination.

The main requirement is the text needs to be long enough… just a small sentence or snippet won’t be enough.

Having said that, it’s not that hard to defeat the watermark once you know how it’s done. If you know it operates on token trigrams, then you need to rewrite the document at the trigram level to break up those relationships.

Are they just trying to check a box on some compliance checklist?

Actually yes, this has been prompted by a new EU law requiring AI companies to make LLM output identifiable so that people have a chance at knowing what is generated/fake content.

This is a futile approach. It has been shown again and again that watermarks can easily be removed by AI models that are widely and openly available. It is like a bandaid for a severed artery and only is useful to shield Anthropic from the wrath of the courts.

That only thing sounds like strong motivation for them.

placebo

I think this is honestly one of those 100% vs >80% issues. I get that it’s very possible — maybe even easy — to subvert this, but we have had several SITTING POLITICIANS read a fucking AI prompt in the middle of their “big speech” and nobody bats an eye. If they add a measure to outputs that works automatically for 100% of outputs without the people reading it knowing, there’s no chance that it won’t allow you to determine whether or not many outputs are AI.