Technical critics argue watermarks are trivial to defeat
3 Aug 13 · 66d ago · 16 articles · 53 posts · 62 comments · 6 sources · development 3 of 8
Software engineer Sean Goedecke published an analysis arguing text AI watermarks will always be easy to remove, and James Padolsey published an interactive essay explaining the statistical word-choice mechanism, both drawing large HN discussions.
Anthropic AI lab, maker of ClaudeJohn Gruber Daring Fireball writerArturo Villarroya Claude Max subscriber, public policy workerVladislav Rajtmajer Freelance developer, Czech RepublicJames Padolsey Author of interactive watermarking explainer
The whole story articlespostscomments the bright band is this development · numbered dots are the others · click one to jump
Reported in the same hours no headline names this development itself — these 10 claims were published in its stretch
-
first by Nature News, 66d ago
-
1 outlet How AI text watermarking works
first by HN Frontpage, 66d ago
-
first by Nature, 66d ago
-
first by HN Frontpage, 66d ago
-
first by Politico Europe, 66d ago
-
first by Business Insider, 66d ago
What people said 24 voices · best of 84 · verbatim
-
0
@theverge.com Anthropic says it will put an invisible watermark on #Claude-generated text and images in an effort to comply with European rules for AI transparency.
-
C
🚨 Acaban de crear una skill que elimina las marcas de agua de Claude. Se llama watermarks-remover y está diseñada para limpiar las diferentes señales invisibles que pueden quedar incorporadas en textos y archivos generados con IA. Esto es lo que hace: → Elimina caracteres Unicode invisibles y espacios especiales → Limpia metadatos C2PA, EXIF y…
-
One aspect that people seemingly aren't talking about is the impact this has in the model's creativity. Because the model will nudge each word towards group A vs group B, you're losing on creativity, especially more so if the nudge isn't a gentle 55% but something like 70% or 80%. So essentially they're forcing the model to be less creative for…
-
Good. If it doesn't catch the person that paraphrased, if it doesn't catch the person that used the AI as a tool and then put their on thoughts in their own result, but if it catches "You" who simply copied and pasted an AI response, then that watermark is doing exactly what it's supposed to do. If you hand in your dissertation, and your advisor…
-
S
If yiu use AI in your job to keep your job, you dont deserve your job.
-
B
“watermark in text” seems so trivial to defeat; text is, in its computer representation, extremely well formed and easy to detect unwanted junk. I've been wanting to respond to the hand-wringing about “oh no, how will I survive if I can't trust the # GenerativeAI output not to have unwanted junk inserted into it to “watermark” it” and, yeah it's…
-
From: https://support.claude.com/en/articles/16266773-how-claude-m...> We'll support users and other third parties to detect Claude’s marksThis 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…
-
Open weight models don't do this and erasing the statistical watermark is not hard and simply requires reprocessing the output with one of those models. It will likely become a common workflow for those who are worried that takes seconds. Once tools are available to identify the watermark, it will become even easier to remove. So this really only…
-
J
what do we think, 3 minutes until someone makes a "claude watermark breaking" tool, "hey claude, break your watermark", thanks.
-
M
With full acknowledgment that a). this is hilarious, and 2). it is good to know what is and is not AI, this gives me an icky feeling. Fools claim they can spot AI based on em dashes. Schools use hugely unreliable AI-detection software. Will this be any better? https:// techcrunch.com/2026/08/12/some…
-
Suppose I gave you a series of 1000 coin flips. I tell you they were generated by a fair coin. You’re suspicious, you think I hid a watermark in them. But you look at the sequence and 485 are heads, close enough to 50%. You look at the correlation between a coin flip and the next and it’s 0.0433. Close to zero. You do a bunch more stats and…
-
No need to give it to a team of people, just pass it through a different model and ask to rewrite while keeping semantic and factual meaning. Heck, even just passing it in a translation service might do it. Ask Claude to generate whatever you need in say French, then go to Google translate and translate to English. Voila, no more watermark.
-
K
Anthropic dice que marcará con una marca de agua el texto generado por sus modelos de IA —
-
S
The watermarks-remover tool combines two different techniques: one removes identifiable bytes (verifiable), the other rewrites text to disrupt statistical patterns (unverifiable without Anthropic's detector). This gap between what can be proven and what is claimed matters for trust in removal tools. https://www. implicator.ai/github-tool-targ…
-
For SynthID and similar solutions, there is much I don't understand ...Here's what I grasp: The AI system scores each token and then selects tokens based on those scores. If we encode something in the token selection routine ('in order choose the 1st, 3rd, 1st, 5th, 2nd, then 1st highest scored tokens'), we can identify AI-generated text by…
-
The quality of information lies not in who wrote it or who communicated it, but in what the information says. A philosopher may have written many ideas by hand, but all of them may be wrong. Whereas a text processed by an LLM may contain true information, relevant arguments, and well-constructed reasoning.
-
H
How AI text watermarking works ---- - 53 minutes ago | 10 points | 3 comments - URL: https:// declaude.org/watermarking/ - Discussions: https:// news.ycombinator.com/item?id=4 9292932 - Summary: This article explains how AI text watermarking works through five steps. Watermarks don't live in characters but in the *choices between words*. At each…
-
The article mentions you can always simply use a smaller, local, un-watermarked LLM to rephrase the original watermarked text. Which is true, sure.But if we're talking about deterministically taking some watermarked LLM output and having a function removeWatermark(text), it won't necessarily be "trivial" to remove, because the watermark function…
-
Most folks I know apply to over 1000 places for each job hop. I’d be anguishing over what words I use where and it takes me longer than most, but even if it’s just a minute that’s still like 16 hours of extra work over 1000 applications.
-
A
# Anthropic announced that its # Claude models will embed an # imperceptiblewatermark into generated # text to comply with the # EUAIAct . This # watermark , likely implemented through # statisticalwatermarking during token generation, aims to survive copying and minor edits. While this move aligns with EU regulations, it could potentially lead to…
-
This is a terrible idea. It sounds like the way they're going to do this is by looking at specific word choices, which is the standard say AI "detection" currently works. The problem is that if it is "undetectable" then it is fully valid language, and if its fully valid language, someone likely talks/writes like that. A lot of my pre AI era work…
-
yes, but the water mark doesn't differentiate between someone who did the work, and used AI to clean it up, vs Someone who told AI to do the work, and didn't even bother to read the output.
-
O
「AIを使った「透かし除去ツール」がウェブ上に溢れている。しかし、実際に効果があることを証明できるものはほとんどない。 」： # BLEEPINGCOMPUTER 「AnthropicがClaudeの著作物すべてに目に見えない透かしを入れることを公表してから数日後、AI透かしを除去する市場が急浮上した。GitHubプロジェクトには4,500以上のスターが集まり、新たに登録されたウェブツール群や、少なくとも1つの既存のAI検出回避サービスも登場している。 しかし、テキスト透かしを無効化できるという主張は、現時点ではどれも検証できない。なぜなら、Anthropic社はまだその仕組みを公開しておらず、クリーニング後の文書に透かしが残っているかどうかを示す検出ツールもリリースしていないからだ。…
-
"What would it even look like to sign ChatGPT outputs? There’s no artifact to pass around."It would look like a lot of little signatures on little bits of text, and then larger signatures on a collection of those chunks once the larger chunk exists. It's not that hard. It's just a lot of signatures.But this scheme could only ever prove that this…
All 8 developments of Claude's AI text watermark sparks backlash, cancellations… →
Hacker NewsNewswiresMastodonRedditXBlueskyGoogle NewsLobsters