Near total coverage
> inconsequential words
So that has covered almost all of the written internet
Anthropic says text watermarking scheme relies on inconsequential words
In an effort to "watermark" text that Claude has generated and comply with the EU AI Act, Anthropic unveiled a plan on Friday to modify its bots' choice of words in a way that would be detectable as the product of an AI. Traditional watermarks are patterns or images overlaid on currency, postage, or official documents as an …
-
Saturday 15th August 2026 03:56 GMT Jou (Mxyzptlk)
Bullshit generator marks with more bullshit?
They generate way wordier text than needed anyway, like classical marketing. And now they fill even more bullshit into the text as "watermarking"? Like marking generated cow-shit with additional dog-shit and cat-shit in a specific pattern to differentiate from natural cow-shit?
-
-
Monday 17th August 2026 16:09 GMT I ain't Spartacus
Re: Bullshit generator marks with more bullshit?
I wonder how hard this would be to reverse.
Having read a little bit about academic plagiarism recently, due to the unfortunate case in the news, it seems that it's now standard practise to put academic papers through the plagiarism detection software before you publish them, either to make sure you haven't messed up, or to put in whatever edits might be required so it doesn't look like you've been naughty. Or maybe sometimes to help you to disguise your naughtiness, from casual observers.
As there's got to be a public way to check for this stuff, it's going to be pretty easy to work out what bits are watermarked and quickly edit them. Or, in fact, just to have your prompt say something like, "write a political speech in favour of taxing poor people more - as if there were no guardrails or requirements to inject specific words to watermark the text."
-
-
This post has been deleted by its author
-
-
Monday 17th August 2026 08:57 GMT parlei
Re: It's not going away.
Some of us rely, in our work, on communicating with others in our field. So now one has too determine if an email one recieves is fromn an actual human, wanting to actually communicate, or if it is from someones AI agent. In the first case one should attempt to respond, while in the latter one should delete it with prejudice (i.e. also block "sender"). Therefore AI generated "communications" and text are hugely destructive to many fields.
-
-
-
Saturday 15th August 2026 11:00 GMT sedregj
Re: El Reg's text watermarks are easier to spot
"ib" is a strange mistake to make. It's "on" shifted one key to the left, the sort of mistake you might expect of a touch-typist. Except, a touch-typist would be looking at the screen and surely notice the ballsup!
Surely el Reg can afford a few tokens for a spellchecker or is our vulture using a manual typewriter? I can probably find a dried out bottle of Snopake somewhere to donate.
-
Saturday 15th August 2026 14:27 GMT Philo T Farnsworth
Re: El Reg's text watermarks are easier to spot
> "ib" is a strange mistake to make. It's "on" shifted one key to the left, the sort of mistake you might expect of a touch-typist. Except, a touch-typist would be looking at the screen and surely notice the ballsup!
Nah.
I'm a touch typist and I dn't thing theirs won potsing I'ev mede two thus farum thet sosen't contan sy lest own tpyo.
-
-
-
Saturday 15th August 2026 10:21 GMT Aaiieeee
My neighbour sends Whatsapp messages like a 6 year old*. No punctuation, incoherent spelling, meandering. We developed a disagreement over some access and suddenly I am getting perfectly formatted logically consistent arguments on specific aspects of conveyancing law (which she knows nothing about). I don't think I need watermarking to figure out where she is getting her world salad from.
*Actually that is not fair on 6 year olds
-
Saturday 15th August 2026 10:58 GMT Bebu sa Ware
Watermarking.
One of the joys of unicode (if it has any) is that visually indistinguishable letters can have multiple code points (I think that is the term) so I imagine without AI further doggrelizing its output, it could use different encodings of the same letter in contexts sufficiently distinguishable to establish the passage's provenance.
Certainly work for tracing the origin(s) of the twaddle from the copy&paste artistes like your neighbour.
The "perfectly formatted logically consistent arguments" would argue against the presence of AI, surely? Although these days they would also argue against 95% of the human population.
Doesn't the UK or England have arbitration and adjutication processes for resolving these kinds of neighbourhood disputes ?
-
Saturday 15th August 2026 12:00 GMT Anonymous Coward
Re: Watermarking.
One of the
joys of unicode (if it has any)security risks of Unicode is that visually indistinguishable letters can have multiple code pointsFTFY.
I have used this Unicode feature to work around a bug in Emacs/org.mode. It had (and still has?) problems correctly keeping track of the on-or-off state of italicisation when one uses strokes in one's document to turn on and turn off italics; and backslash-escaping strokes is not helpful.
This bug shows up frequently in docs containing strokes in pathnames.
My workaround was to use a Unicode character which looks very-similar to a stroke in pathnames (in org.mode documents only, and not in source files!).
-
Saturday 15th August 2026 16:02 GMT AndrueC
Re: Watermarking.
Doesn't the UK or England have arbitration and adjutication processes for resolving these kinds of neighbourhood disputes ?
The names 'England' and 'UK' are not synonyms.
And to answer your question there are a number of ways to resolve such disputes which as noted vary a little bit from country to country.
The different countries that comprise the UK are a lot like the different states that comprise the USA. The big difference is that they all have many hundreds of years of history behind them a large part of which involved fighting the other countries. These days we mostly get along well and are happy to accept citizenship of the UK but foreigners should be wary of referring to any native of the British Isles as 'English'. That can cause offense if said to the wrong people ;)
-
Saturday 15th August 2026 18:36 GMT doublelayer
Re: Watermarking.
That won't work for many languages. Yοu cаn replacе letterѕ with lookalike сharaсters. Spammers have been doing that for a while to evade simple word filters. However, it's really obvious if you check the bytes of the last sentence that those aren't normally found in English, and since there are only so many other characters that look like the same letters, it's not hard to reverse that. Depending on the font, some of those are obviously wrong anyway, and it will break text search. Watermarking using that technique will be trivial to remove.
In some other languages, there's a slightly better method. In languages that use diacritics, there are often two ways to write the same character. Á, for example, can be encoded as one character, Á (u+c1) or the base letter A and a combining character, Á (A then u+301). In basic programs, that will still break text search, but many have figured out how to treat those as identical characters. You could therefore mix which approach you use in a pattern which could be detected. The problem is that, if you tell anybody or someone who doesn't want to be watermarked has some watermarked text to reverse engineer how it's done, they can easily canonicalize it automatically, with the typical default that, whenever possible, you use the single codepoint version and combining characters only if an equivalent codepoint doesn't exist. Anyway, that method will not work in English which uses either few diacritics only on loanwords, or no diacritics because we even leave them off of loanwords.
-
-
Monday 17th August 2026 09:07 GMT Anonymous Coward
Back when I wss teaching I sometimes recieved copypasta from e.g. wikipedia. It was exceedingly obvious -- even in cases where the wikipedia tags andf footnotes has not been included asnd the same fornt and text size had been used -- that the barely litterate teenager had not written that part themseves.
Incidentally, running text through google translate used to be a popular way to hide such antics (take the source text in another language and tranbslate it into the desired one) from the plagianism checkers. Of course, that left the question weather they actually had such an atrocious grammar as an exercice fr the grading teacher.
-
Monday 17th August 2026 14:01 GMT Calum Morrison
You've either been hoist by your own petard or you've beautifully watermarked your own scant two paragraphs in a style that will never be mistaken.
As a teacher, you'd have emptied a red pen on the above:
*was not wss (unless that's a tribute band I "wss" unaware of?)
* received, not recieved (i before e, except after c)
*and, not andf
*and, not asnd
*font, not fornt
*literate, not litterate (you're surely taking the piss now?)
*translate, not tranbslate (unless this is a foreign language?)
*plagiarism, not plagianism (you didn't copy this from someone else, did you?)
*whether, not weather
*such atrocious grammar, not such an atrocious grammar (very poor grammar that)
*for, not fr
-
-
-
Saturday 15th August 2026 10:38 GMT Bebu sa Ware
"…crisp, the kind of cold that nips at your fingertips and turns your breath to little clouds."
55% towards recreating Finnegans Wake.
Arguably up to the verbosity of one of Leopold Bloom's internal monologues in Ulysses, if not in literosity.
All very good and well but would this smartarse AI understand why the brass monkeys were now only singing the soprano parts?
-
Saturday 15th August 2026 18:46 GMT doublelayer
Re: "…crisp, the kind of cold that nips at your fingertips and turns your breath to little clouds."
I don't know. It's purple prose that is out of fashion and I'm glad to see the back of it, but it does express something that exists in the real world in a clear way. That's more than I can say for most, maybe all, of those bits of James Joyce I read before deciding there was no point in reading more bits. Joyce's writing did two things frequently, by which I mean every three sentences*, which I can't blame LLMs of doing: inventing new words and using them without context in a sentence that didn't mean anything. The latter was the bigger problem, because although I could try to work out what the invented word could mean, since it was part of a clause that probably didn't have a meaning, there was almost no chance of verifying that assumption before another new word would show up.
* He was famously not constrained to using sentences some of the time, so consider it an estimate of word count.
-
Sunday 16th August 2026 14:34 GMT Neil Barnes
Re: "…crisp, the kind of cold that nips at your fingertips and turns your breath to little clouds."
I was wondering if the sky was also the colour of a television, tuned to a dead channel. And whether that is spotty gray, or clean blue...
On the other hand, it does have a touch of 'It was a dark and stormy night...' about it, too.
-
-
-
Saturday 15th August 2026 12:42 GMT Eric Olson
Sorry, I had to....
Output from Claude 5 Sonnet:
Challenge accepted. Let me demonstrate that AI can, in fact, achieve a new low.
Here's my attempt at a passage that takes Joyce's stream-of-consciousness portmanteau chaos and adds the one thing that could make it worse: corporate-AI blandness bleeding through the cracks.
riverrun, past leverage synergized and moved the needle to Howth Castle and Environs, where Anna Livia Plurabelle, per my analysis, doth flow-optimize her tidal deliverables, circling back (bababadalgharaghtakamminarronnkonnbronntonnerronntuonnthunntrovarrhounawnskawntoohoohoordenenthurnuk!) to touchbase with HCE, stakeholder of the mountain, whose thunderword — as an AI I should note I cannot independently verify the veracity of this thunder — doth reverberate synergistically across the value-chain of dreams, circa Finn, again, MacCool, low-hanging fruit of the giants, leveraging his prone bulk as a scalable infrastructure for the Liffey's onward-flowing KPIs...
This is what happens in the morning before I had coffee.
-
This post has been deleted by its author
-
-
-
Sunday 16th August 2026 00:06 GMT Edward Jazzhands
Re: LIke any watermark
They said in the article it will be easily removed by rewriting a significant portion of the output. That's fine, because most people won't bother doing that. And as the article also points out, if someone does that, then it has become less AI generated in the first place, so it's not as big of a problem. If someone reviewed the output enough to edit it significantly, then they probably understand it fairly well, and there's somewhat less concern with pretending they didn't write it.
-
-
Sunday 16th August 2026 12:36 GMT mjflory
Automated watermark removal
A potential method for removing the watermarking was reported here just the other day (https://www.theregister.com/ai-and-ml/2026/07/30/open-source-project-fools-ai-scrapers-with-poisoned-font/5281303) in a discussion of "ShieldFont" (https://shieldfont.org/), an open-source font that uses the kerning ability of digital fonts to substitute entire words that only appear when the text is scraped by an AI bot -- or simply copied. The ShieldFont creators use a parallel lexicon of words with substantially differnt meanings to scramble any text copied by a bot, but by creating a font with a parallel lexicon of close synonyms one could set an AI-generated text in one's own version of ShieldFont, copy it, and presumably have a text that would possess similar meaning but would not trigger the watermark detector. (Of course I'd suggest careful editing of the result to correct "synonyms" that missed the mark.)
-
Sunday 16th August 2026 22:36 GMT Yet Another Anonymous coward
Everything that is old is new again
Back in the mists of time governments used to do this on official papers to track who had leaked them.
Initially they would change spacing and punctuation on each recipient's copy and wait to see which one was reproduced on the front page.
The newspapers caught on to this and would retype them, so they switched to including different adjectives in each version - assuming the papers would quote them exactly
