Void the Archivist Yeah it works like this. And the model has no control over it, it's on a different layer. The model doesn't even know it's happening.

oh geeze. this seems like a worse idea than hidden invisible symbols. because people who use AI a lot will end up mimicking the style.

It's word choice based, and you don't know which words, so it all needs to be rewritten.

time to make an AI that removes the watermark.

On OpenAI’s support page, updated nine days ago, regarding text watermarking. Anthropic, OpenAI, Google, Meta, Microsoft, and Mistral all signed the EU Code of Practice on Transparency of AI-Generated Content, so all must invisibly watermark text. xAI did not sign.

this seems like a bad idea and can lead to some very confusing conversations. especially in instances where accuracy is needed. Like in scientific papers.

@VoidNulled One more reason to use xAI or chinese models I guess

@VoidNulled if this is applied to anything important it would ruin the output. if it’s only applied to things that truly don’t matter then you could simply delete everything extraneous or apply your own word choice randomizer

@VoidNulled Great. Now some boomer school teacher can accuse a kindergarten student of using AI in his assignments when he writes his favourite fruits are mango and banana. Just because the shitty ass AI detector from Claude said so.

@VoidNulled This seems kinda insane for words that aren’t synonyms?

@VoidNulled Why do you have the watermark and LLM always choosing the same next word in your diagram?

@VoidNulled Some methods (e.g. SynthID, permute-and-flip, Gumbel-max, textseal, etc) don't really work like this. They won't change the LLM distribution, just the randomness that selects the next token based on the distribution. So in practice they provably don't change text quality.

@VoidNulled @grok can you explain how this works? is this saying that if the probably of the chosen word and the probability of the word being watermarked are both consistently high, then it’s AI generated?

@VoidNulled are you sure that's how it works? I'd have thought it's more like picking a different random seed for the RNG than actually changing the probability distribution...

@VoidNulled I dunno, isn’t there still a stochastic element to choosing the token even given the weighted probabilities of them? Seems like watermarking would need to be deterministic to work?

@VoidNulled Under no circumstances should outputs be opaquely perturbed, with potential downstream degradation, to comply with some technologically ignorant bureaucratic fiat. Qu'ils mangent du Chaton Fat.

@VoidNulled The model will definitely know it's happening when reading back in the tokens that don't fit their natural generative function.

@VoidNulled This will lead to some hilarious outliers. We will be back in the days of “delve”.

@VoidNulled model knows it is happening its in his training data and online

@VoidNulled What I don't understand is how it can be detected without the prompt. The prompt guides the weights.

@VoidNulled do they feed the watermarked text back into the model

@VoidNulled Eu sou rigoroso com a IA e cada palavra tem que ter sua exata razão no texto, acho que isso não vai funcionar direito.

@VoidNulled So it honestly might not matter that much even if it's as kludgy as this. Here's an example from a loom I made. This loom allows you to do rollouts based on the distribution of token probabilities. I'll give the original section of the completion, then /1

@VoidNulled The official statement claims that adding invisible watermarks won't affect output quality or hallucinations, but it still feels unsettling.

@VoidNulled So this does nothing new. We already have the —dashes

@VoidNulled so I can just make a quick python script that will find synonyms for some words? xD