How about we just… turn off all the chat bots, and toss the whole ball of garbage on the trash heap of history?

I don’t see any point in wasting power on any of this, much less the additional burden of watermarking.

None of it can be trusted, and it actively harms us at both an individual level, and a societal one.

…I’m so tired of this timeline.

Like the recycle mark, it will likely be an industry strategy to put it on everything and make it so confusing that it really doesn't convey any information at all (positive or negative) and becomes ignored. Mission accomplished. Regulations gamed. Actions performative. Costs, practically zero.

I am curious: how do you watermark ASCII?

It's in the article, but for text at least, subtle word choices that make a pattern across the whole document.

Makes sense to me to put it on everything. LLMs don’t do simple grammar fixes, they just partially rewrite passages in their own voice. They’re not deterministic the the old grammar checkers were. Lots of potential for them to hallucinate, especially by tweaking meanings at a subtle level. I’d love a way to tell that someone had partially sloppified their writing.

I imagine this is going to be something that is trivial for a machine to implement but difficult for a human to detect such as making every 17th word end in "-ly". Thats why in some cases the word chosen may not be the ideal word for that sentence but it serves to maintain the streak pattern.
It would resist mild editing because removing 1 or 3 of those "-ly"s would preserve the rest so that if 90% of every 17th word ends in "-ly" the statistical odds of that being just a coincidence is nill and the watermark is preserved.

The actual technique is a bit more involved than that. It basically uses something equivalent to a secret key to influence generated output that do not make a significant difference in meaning. For example, if writing a story about a dog, and it has a choice of 3 colors to choose from to describe the fur color, and another 3 words to describe the length, it will use the private key to select the output in a pseudorandom manner. Given enough of these across the document, you can re calculate the key used to generate the article, even if it is cut up, moved around, or edited in some spots. The signature will be nearly impossible to detect without knowing the key used to generate the output.

In the novel, the scarlet letter is extremely visible - that’s the point. It’s also a mark of shame…ok, now the name makes sense.

All of these techbros are functionally illiterate. Everything they do has some literary allusion that's supposed to be clever but actually just misses the point entirely

How about we just… turn off all the chat bots, and toss the whole ball of garbage on the trash heap of history?

I don’t see any point in wasting power on any of this, much less the additional burden of watermarking.

None of it can be trusted, and it actively harms us at both an individual level, and a societal one.

This is the standard refrain with new technology. The problem is that some abuse it and make a nuisance with it. However, there are people getting actual, reliable work done with it.

Throwing it away is not the answer, much like it wasn't the answer for any technology that came before it that people got frustrated with because everyone got carried away with them and dishonest people tried to use to their advantage.

@Aurich The choice of image was already great. The fact that their logo is clearly a butthole gives it that final je ne sais quoi. A+

There are a surprising amount of AI companies whose logos are basically buttholes.

I tried using Claude for proofreading and frankly it sucked. It messed up on many comma placement issues. I ended up finding grammarly worked more consistently. It was weird in that claude seemed to work, but it would just skip errors. It probably caught 80% I'm not really sure. Enough that I thought neato it works, then I read over it and thought, hmm it missed this.

This was the prompt I gave it, perhaps it was too complex. Some of the things in it are because it specifically missed those errors the first time I tried it, for example it seemed to skip any error in a sentence if it already found one. Maybe they have improved since I tried it but I doubt I'll try again since grammarly worked better anyway.

Your strict constraints: 1. DO NOT change the style, prose, or flow. 2. DO NOT rewrite paragraphs to make them sound smoother. Your ONLY jobs are to look for and flag: 1. Mechanical errors — including typos, missing words, missing or misplaced punctuation required by standard grammar rules (such as commas around direct address, before coordinating conjunctions joining independent clauses, after introductory clauses, around similes, and around appositives), missing quotation marks or periods, and run-on sentences. Do NOT flag punctuation omissions that appear to be deliberate stylistic rhythm choices.

DO flag them in sentences where the missing punctuation has a clear grammatical function. 2. Structural duplicates (cases where I wrote the same thought or sentence twice in a row in different ways because I forgot to delete the draft version). For every sentence you flag, re-examine it for any additional mechanical errors before moving on. Format your output as a clear list: - [Location]: (Paragraph number or starting phrase) - [Issue]: (Describe the duplicate sentence or clear typo). If a section has no mechanical errors or duplicates, simply reply with "No errors found." Analyze the following text:

The mark flags anything Claude processed, even human writing it only edited.

Not sure if AI watermark or human just.

I thought we already had text watermarks, isn’t it the em dash — ?

1. Emdashes were never perfect slop signifiers, since plenty of literate and/or typographically aware humans used them (which is where the bots learned about them in the first place.)

2. The slop slingers have been adjusting their base prompts to cut down on emdashes

3. Perhaps reflecting the first two points, The Economist recently did a textual analysis of slop vs. articles written pre-AI, and found that emdashes weren't actually that much more common in slop. Their biggest finding was that bots tended toward grandiose prose, which I can definitely believe, since I pretty much always have to include some variation of "be concise" in my prompts.

On top of this, you have the general public, who may not grasp the difference between processed text and wholly generated text. If the system watermarks human-authored text simply because it was edited in a workflow that touches Claude, suddenly it carries the same denotation as wholly generated text does.

As an AI scientist, I think this is fine. The difference between AI-processed text and wholly generated text is irrelevant in practice. Both are compromised and not worth reading.

This is 100% malicious compliance ment to muddy the waters.

Are you sure? It might be just a PR spin on watermarks that they need for their own use - to prevent their own slop from becoming a part of training set.

What's the minimum length to actually create a watermark in ascii? If it's a bot writing comment spam in a specified style, a couple of sentences don't leave a lot of room for a mark.

The word-choice technique is a lot easier if you have both the prompt and the generated text. Without the prompt, you can't know whether the word choice was directed by the user.

Also, will this affect direct quotes? Like if I ask it for a passage from Shakespeare or the U.S. Constitution, and it sprinkles in "similar" words, that's going to cause problems.

Feel like by ensuring that it's used everywhere/on any touchpoint, their hope is it'll become pointless?

It's already pointless. I think their goal is just the easiest form of compliance that will keep them safe from regulatory action.

So now if I ask it to perform a simple spellcheck it's going to suggests some useless word substitutions just to create a watermark? Sounds great...

Considering how bad normal autocorrect is getting, would anyone notice?

There may be a self-serving purpose; the internet is starting to accumulate a lot of AI slop, and including it in training materials can undermine AI models. (Model collapse.) If the companies mark AI outputs in some way, it could protect their ability to maintain the models.

If AI content is so great, and full of so many wonderful new ideas, why wouldn't AI corporations want to train their own models on their own output?

I hate the EU sometimes. They should focus on out-competing American tech with sovereign tech. They should give a reference implementation to help AI startups do the watermarking. Instead, they pass expensive regulations that only the largest companies can comply with, in an opaque manner, handing Big Tech even more power through regulatory capture.

As another example of stupid EU regulations: the EU Digital ID system requires hardware attestation which hands iOS and Android even more power and locks out open source alternatives like Graphene.

I feel like running these through another AI could "demark" it, but as of yet, the fakery is still detectable by human eyes, just not immediately in some cases. That, and when it comes to deep faked major events, the same rule applies: if the image is the ONLY shot of something major like all of Hollywood burning down, it's probably a fake. If the only place you saw it was a captioned picture without a linked source on The Facebook, it's probably a fake. As of yet, generating two unique camera angles of the same scene is still pretty inconsistent for these LLM generators to handle.

There's something else though, and that's how upset corporations are at having to be up front and honest about it. If you truly believe in the tech, sing it with your full throat. If you feel the need to "hide it" in order to sell it to people, what does that tell you about public acceptance?

Well, then Claude will be useless to me, which might be the point. I am using the project feature to host the A-Z of books that i am writing. That document alone is over a hundred pages, scenes, characters, the world-building, the blindsides, basically everything about the books without them being fully written yet. It has taken me a year to write that document so far and i guess thousands of edits, and i am not even sure when that part will be "finished" if ever.

The writing itself, i do, but i do ask for wording suggestions as I am not a native English/US speaker. Time to bring it locally i guess.