The most prestigious law school admissions discussion board in the world.
|
 |
 
|
New Claude models to include watermarks inside text - link
Poast new message in this thread
Date: August 10th, 2026 5:29 PM Author: German pumo https://x.com/M1Astra/status/2086898041882030353
Does this mean teachers, judges, etc will be able to know you used Claude AI? Seems like a gift for ChatGPT if so. |
Date: August 10th, 2026 5:31 PM Author: The Penis Open AI already looked into doing the same thing and decided there were major drawbacks (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058695) |
 |
Date: August 10th, 2026 5:37 PM Author: https://i.imgur.com/wa2B8mT.png (gunneratttt) It seems useless anyway since you'd only have to change a few words and the only value is AI companies and governments acting like they're regulating AI transparency. (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058710) |
|
Date: August 10th, 2026 5:33 PM Author: redpilled tradcath nietzschean eugenicist girldad idgi. cant you just be like claude write me a 500 wd essay, then print it out and retype it into word? So you're just adding one step.
tbf even adding that one step might actually cut down significantly on ai plaigiarism. idk. people react funny to things |
|
Date: August 10th, 2026 5:34 PM Author: The Penis No the "watermark" isn't in metadata or characters. It's a token bias that causes certain tokens to be selected over others in a way that doesn't change the overall meaning of the text, but it can be recovered that it was made by claude. You can still destroy it by rewriting pieces of the output or running it through a translator and back.
One of the reasons Open AI decided not to do this is they thought it could cause a DITHPARATE IMPACT to ESL users |
|
Date: August 10th, 2026 5:36 PM Author: nietzschean girldad ogling his own teen daughters it's done via semantic patterns woven into the textual language itself (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058707) |
|
Date: August 10th, 2026 5:38 PM Author: redpilled tradcath nietzschean eugenicist girldad wowza
like how shakespear supposedly snuck his name into Psalm 46 |
|
Date: August 10th, 2026 5:42 PM Author: the wife guy cinematic universe Yeah I worked on something like this in college for a corporate espionage thing (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058720) |
|
Date: August 10th, 2026 5:33 PM Author: Nazca Redlines How would this work? If it's part of the text, the user can just delete it--assuming they can be bothered to look through the text to find it. (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058700) |
|
Date: August 10th, 2026 5:36 PM Author: OYT was Right All Along ( ) "it will travel with the text when it's copied and pasted elsewhere, and may persist through some editing."
they know it can just be deleted, but many people are dumb enough to do a blatant copypasta |
|
Date: August 10th, 2026 5:40 PM Author: the wife guy cinematic universe You could watermark AI text by using a secret key to bias otherwise natural word choices.
At selected points, a keyed PRNG decides which of several semantically equivalent choices is preferred. Those choices encode bits, and a verifier with the same key checks whether the observed choices correlate with the expected pattern.
For example:
"however" = 0, "but" = 1 "large" = 0, "big" = 1 "because" = 0, "since" = 1
If the secret pattern is:
1 0 1 1 0 0 1...
the model subtly favors the corresponding choices. To a reader, the prose still looks completely normal.
A more sophisticated version would not use obvious synonym pairs. Instead, candidate next tokens could be secretly divided into two groups with something like:
bit(token) = HMAC(K, context || token) mod 2
To encode a 0, slightly favor tokens from group 0; to encode a 1, favor group 1. The result is a keyed statistical signature rather than an obvious hidden message.
You also would not need every bit to survive. Add redundancy with an error-correcting code, parity bits, and repeated/interleaved copies of the watermark throughout the text. Then even if large sections are deleted, reordered, or rewritten, the verifier can recover enough surviving signal to identify the watermark statistically.
For example, across 500 surviving tokens, if 68 percent match the secret expected groups instead of the roughly 50 percent expected by chance, the verifier can assign a confidence score that the text came from the watermarked generator.
|
|
Date: August 10th, 2026 5:45 PM Author: The Penis Yeah its retarded though because you can technically destroy it just by changing one token, because that can alter the watermark state for subsequent positions. Theoretically, depending on the specific strategy they are using for biasing it, and which token you pick. But the more robust they try to make the watermarking, the more constrained and janky the writing and output capability becomes. (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058734) |
|
Date: August 10th, 2026 5:43 PM Author: https://i.imgur.com/wa2B8mT.png (gunneratttt) A pattern is encrypyed into the text. Think something like making every word occurring in a prime number sequence starting with an advancing vowel, but obviously more complicated. An AI could write anything under those rules without changing the content or quality that much. But yeah, obviously anything encrypted into plain text can be destroyed by simply changing a couple words. (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058724) |
|
Date: August 10th, 2026 5:39 PM Author: queensbridge benzo taking a screenshot then telling chatgpt to type the text from it is an obvious workaround (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058714) |
|
Date: August 10th, 2026 5:46 PM Author: The Penis Huh? How would that do anything? (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058739) |
|
Date: August 10th, 2026 5:44 PM Author: https://i.imgur.com/wa2B8mT.png (gunneratttt) AI writing "I want to kill myself too" next to your 4pt white text watermark. (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058727) |
|
Date: August 10th, 2026 5:49 PM Author: botman "I'm gay"
- This text was generated using Claude Fable 5, a "Mythos-level" frontier AI model by Anthropic. |
|
Date: August 10th, 2026 5:52 PM Author: nietzschean girldad ogling his own teen daughters (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058754) |
|
Date: August 10th, 2026 5:52 PM Author: suicide hotline bling (Prestige Feuds) (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058756) |
|
Date: August 10th, 2026 5:58 PM Author: redpilled tradcath nietzschean eugenicist girldad (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058771) |
|
Date: August 10th, 2026 5:59 PM Author: shitlaw boss vibecoding productivity tracker (🧐) (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058776) |
|
Date: August 10th, 2026 6:02 PM Author: The Penis (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058782) |
|
Date: August 10th, 2026 6:03 PM Author: taco bell nigga Write your own words you weird fags (http://www.autoadmit.com/thread.php?thread_id=5891530&forum_id=2]#50058785) |