Claude to watermark all text generation because of EU laws. This is for censorship/control.

[CaptainTrouble]

Now, if you use Claude to generate text and copy+paste it anywhere, the EU knows it's you (since you need an account that you likely pay for to use Claude). You can't use your personal info to generate the text and then use an anonymous account to post it. You'll always be tagged on all text generated by the AI. If your text is copy+pasted on a public site and other people copy+paste it then the watermark stays and the EU will know you originally generated the text. If you used AI to come up with a popular phrase like: "To know who rules over you, look to whom you cannot criticize" and a bunch of White Supremacist websites copy+paste this text, the AI will know it's you who generated it.

The way it works is by slightly altering the preferred text generative output in a manner that codes the text then Anthropic and decode the watermark but you wouldn't be able to necessarily decode it because their code isn't open source so you don't know the algorithm they use to encode the watermark. To bypass this you'd need to copy+paste it into another AI and have them reword it or you'd need to reword it yourself sufficiently to break the watermark.

There is no good reason for this for the public whatsoever. This will be used to track people down and put them in jail for having their AI generate wrong-think. It also might lead to copywriting of AI generating text or the attempt at doing so such that you don't "own" anything you generate and if you build something (code or a book) using generative text, you may be required to pay to use it if it becomes profitable, etc...

The EU is busy trying to destroy AI with regulations.

How Claude marks AI-generated content | Claude Help Center

support.claude.com

 

[FreeJames]

They can't watermark text. A picture they could. But then people will just use other local AI to remove it, so it won't work.


Edit: ok this is more retarded than I thought. They are planning for large text outputs to map to some value they store. So the order and words chosen will map to a value. Apparently a change of a few words here and there isn't enough to remove the statistical match unless a lot of the text output is changed.

Well that is still easily defeated. But what a bunch of morons for even doing this. It's actually more the government who are the fuckers than the Anthropic.

 

[CaptainTrouble]

No, they can.

The way it works is like this:

Large Language Models (LLMs) generate text one word (or "token") at a time by picking from a list of probable next words. Text watermarking works primarily through Statistical / Token-Level Watermarking:

Word Selection Patterns ("Greenlist / Redlist"):

When generating a response, a secret mathematical key randomly splits possible next words into allowed ("green") words and discouraged ("red") words.

The model naturally prefers picking words from the "green" list.
To a human reader, the text looks completely natural, fluent, and unedited because both options are valid synonyms or grammatical choices.

Detection:

A detection algorithm uses the same secret key to check the text. If an unusually high percentage of words match the "green" pattern, it statistically proves the text was generated by that specific model.


To put into layman terms. They have the output generate an output that it wouldn't normally generate but is still an acceptable output. The variation is the watermark.

 

[FreeJames]

Ya, I got it. Still easily defeated. Put the output into your own local AI and tell it to change it around. Problem solved.

But in reality you only would need to do that in extremely sensitive situations when you need plausible deniability. Or just don't use AI in those situations.

 

[CaptainTrouble]

@FreeJames still, the posturing of the EU to go to these lengths is insane.

This is how it'll be bypassed. Local AI to rewrite the text. Making more work for everyone for no reason. All government is good at.

 

[border_humper]

@FreeJames The revolution will not be sparked by AI generated rhetoric.

 

[Nuremburger]

Seems like a more pro-human move than not, even if they don’t see it that way. This shit is going to be the end of us.

 

[JuliusEbola]

No shit. If a "dissident" can't be bothered to pamphleteer without using AI, they aren't a thought leader of much value as far as I'm concerned.

 

[TOPDAWG]

no its not anyone who thinks that is retarded. you're buying the over hyped BS.

 

[Nuremburger]

I wonder if you’ll say the same thing when the newest data center in town moves next door to you and hums you awake every night? Maybe when the government starts docking your UBI for jerking off too much? Maybe when you’re arrested for your next rolling stop?

We didn’t need this shit before 2022 and we need it less now.

Surveillance state incoming but you retards will bow down and take the chip so long as ‘daddy Elon’ is #owningthelibs!!!

 

[CaptainTrouble]

Government control and regulation of AI will be the end of us, yes. AI without government involvement is/will be extremely beneficial for us.

 

[Nuremburger]

Alright, tell me that in 2035 when we’ve all lost our jobs and you’re begging for your monthly dog food ration in the ‘Universal High Income’ lineup. This isn’t a ‘golden age’ for anyone except the ultra-rich cannibals and pedophiles.

YOU are on the precipice of the end of purpose, skill and prosperity and you’re too ignorant to turn back.

 

[CaptainTrouble]

@Nuremburger Loss of jobs wouldn't be a problem if we deported all non-Whites. Again, a government problem.

 

[ovarieasy]

I still cut letters out of magazines for my ransom notes, so no problem

 

[ChevChelios]

They say that wearing pantyhose over your head defeats Flock cameras as well

 

[border_humper]

where are you sourcing magazines, @ovarieasy?

 

[ovarieasy]

Grocery checkout. I need to know what Taylor Swift and Martha Stewart are up to.

 

[adragons]

B‌e‌c‌a‌u‌s‌e ‌o‌f‌ h‌o‌w t‌h‌e‌y w‌o‌r‌d‌e‌d i‌t t‌h‌e w‌a‌t‌e‌r‌m‌a‌rks‌ ‌a‌r‌e‌ likely invisible characters which are easy to programmatically remove. ‌‌