Live data from Hacker News

Claude users are mad that Anthropic's new watermarks will catch them using it

techcrunch.com

61–70 of 99 posts

Re: Claude users are mad that Anthropic's new watermarks will catch them using it

#61
Based on Hacker News, other parts of the internet and even some job experience, I've come to the realization that many people legitimately have no idea that their pants are down and everything is on full display.

If you thought you could have AI "reorganize a paragraph" and not have anyone know, you're probably wrong. You can't because Claude can't fucking write even a paragraph without ruining it half the time. But apparently, a lot of people are either literally blind to writing style, or just don't care, because even most "Show HN" posts these days are extremely clearly AI slop with little care, marked and sometimes not marked. If you are fortunate enough to be blind to this (trust me it's not particularly fun, the style is truly grating) next time you're on a cool new project, just search/grep for "load bearing" and have a good laugh. or cry.

Of course, not everyone is necessarily trying to be misleading, but:

- I find this writing style to be truly miserable. It's like equivalently annoying to if the entire world started talking and typing in a slick quick-witted cynical way like we were all living in a Marvel movie script or something. Not exactly like that, but it's an analogy. I now realize some people may not notice if they got swapped into such a world and their parents and friends all just started talking like this out of nowhere.

- I find it disappointing when people don't seem to think they could do better, and would prefer to have an LLM speak in their own voice rather than them. It almost makes me wonder if people are even still alive behind the keyboard. Am I utterly insane or is it weird that people are deciding to have something else speak in their voice? For anything short of a speech or PR it just feels weirdly inappropriate. Even if your English isn't great, I probably prefer your English over Claude's. At worst... I'd probably prefer its direct translation attempt over its own attempt to write something.

I've been having a great time with LLMs but seeing the way people use them has introduced me to the fact that my taste is a lot higher above average than I realized. I'm not really thrilled about that realization, because all it really means is the bar is a lot lower than I thought it was, not that I'm going to win an award for having good taste. (It's not that good anyways. It's just that the depths you can go are apparently deeper than I thought.)

Whenever people respond with "well how do I know you're not just dumping out AI outputs" I just sigh. I've definitely seen evidence that people will accuse things of being AI output when they're not, yet I am completely unworried about either false negatives or false positives. If you really can't tell the difference between "the load-bearing smoking gun" and natural human prose, its embarassing for you, not me.

LLMs in this way deepen my already existing dispair, because they're yet again increasing the disconnect between me and the rest of the world.

Re: Claude users are mad that Anthropic's new watermarks will catch them using it

#62

Earlier quoted context omitted.

Unfortunately the witch hunts don't stop whether you use AI or not. Even if you don't use it, (unavoidable) false positives mean the anti-ai crusaders come for you eventually anyway. Happened in plenty of purity-seeking art communities already, will happen in code communities as well. The people who get hurt are the people just trying to make things. Best not to empower self righteous people with flawed sin-detecting…

Agreed... How did the art community deal with the witch hunts? In my case the consequences are mounting. I've already essentially abandoned two communities over it. I'm sitting on human reviewed and tested mesa patches I can't bring myself to submit to upstream because of the AI stigma. I was using AI to revive some of my GCC work as well but abandoned it after their policy announcement. There seems to be no end in s…

It was bad in one that I was part of. First they had people put their work through those ai detectors and there were a bunch of false positives, so they started requiring “process videos” showing you actually working on the project. But then people started calling those AI generated. Nothing could satisfy them. So you’d have artists making handmade work, putting it through ai detectors, submitting process videos of them literally working on the piece, and still they were driven from the community because enough people just got toxic about it. Felt very Red Scare. “Are you now, or have you ever been, a user of AI?"

As an aside, I’ve never seen any pro ai people treat anyone not using ai like this.

Re: Claude users are mad that Anthropic's new watermarks will catch them using it

#63

Anthropic seems to be on a mission to drive away as many users as possible from Claude

Yeah, one would think they'd stop their nonsense given how much competition they're getting. Thank god I've already switched away from Anthropic.

Re: Claude users are mad that Anthropic's new watermarks will catch them using it

#64
From: https://support.claude.com/en/articles/16266773-how-claude-m...

> We'll support users and other third parties to detect Claude’s marks

This is the part that I fear most. What happens with false positives? Will I need to change arbitrary parts of my text if they are flagged as AI generated? Will I be forced to add mistakes?

And then, since LLMs are trained on existing, human-written, quality content, it may be that well-written text gets accidentally flagged more often?

And, if all the providers do this, and they use different watermarks, will I be forced to change things at random until I get a negative from all of those?

I very much think this will yield the wrong signal. There were already some academic plagiarism checkers around that would return false positives, and were blindly trusted by teachers and professors. I suppose AI checkers are all around now, and there will be more of those.

Proofreading is one of the best uses of LLMs that I can think about - I still write my text, then an LLM fixes some of my "errors" since I'm not a native speaker. It usually improves clarity. But now I risk my text to be classified as AI written with potentially bad outcomes.

I don't know whether I should blame the EU, though :-)

Re: Claude users are mad that Anthropic's new watermarks will catch them using it

#65

From: https://support.claude.com/en/articles/16266773-how-claude-m... > We'll support users and other third parties to detect Claude’s marks This is the part that I fear most. What happens with false positives? Will I need to change arbitrary parts of my text if they are flagged as AI generated? Will I be forced to add mistakes? And then, since LLMs are trained on existing, human-written, quality content, it may be t…

Third-party “AI checkers” and LLM watermarking at generation are a bit different. I’d suggest you read some of the methods; they’re approachable to someone on this forum. False positives aren’t really a failure case with good params. False negatives, sure.

Re: Claude users are mad that Anthropic's new watermarks will catch them using it

#67
post #41

Earlier quoted context omitted.

> So we finally tackle the issue of ai-text pollution and probably found a way to clean up the internet We are very, very far from that! So far we have a single ai vendor introducing a statistical bias to their generation that can make it simpler to identify genAI in some (longer) texts. We don’t know yet how effective that will be in practice, and if other models will follow suit

If other AI vendors don't follow suit, then people will move over to them to escape detection. Then of course Claude will stop watermarking.

Watermarking is probably done by other vendors covertly as well. See my other comment.

Re: Claude users are mad that Anthropic's new watermarks will catch them using it

#68
post #8

Not sure why Anthropic is getting all the attention here. Google been doing this since at least late 2024[1] and openai I believe is doing this around 9 days ago[2]. [1] https://www.nature.com/articles/s41586-024-08025-4 [2] https://help.openai.com/en/articles/8912793-provenance-signa...

That’s what I can’t put my finger on, see my other comment : why would Anthropic tell the public.

Re: Claude users are mad that Anthropic's new watermarks will catch them using it

#70

So, it seems the watermark is less what it sounds like (a stamp) and more an "imperceptible statistical pattern woven into the choice of words and sentence structures." Does that mean AI responses will sound even more "AI?" Like, will it become even easier to detect on a read-through because of the word choices and patterning? I see the word "imperceptible" there, but what does this mean in this context? My non-tech…

Even more em-dash.
Post reply on HN