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.
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.
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.
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).
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.
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.
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: https://www.explainx.ai/blog/anthropic-claude-invisible-watermarks-c2pa-august-2026
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
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?
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.
The words and the text itself are meaningless, it’s the content that matters
Removed by mod
placebo
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)…
This should be legally mandatory by use in countries.
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.
