Don’t confuse me with facts! (And, thanks… I have since seen the articles about how it works based on word selection from a subset of the words that would have been randomly selected by the “heat” of the model.)
I don’t know much about it but it sounds like it would only work if you have the weights and maybe the running conditions of the model. Otherwise the statistical wright wouldn’t be known.
It also sounds really easy to break by changing random words.
If you dig a little you can get the articles explaining better, but basically: yes, you have to have the weights, and a secret key… Given those, they split the words that might be chosen at each point into two groups and only choose from one group.
Later, you can check for the watermark by splitting the same way and seeing how many times the words in the text match the group that fits with the secret key. Normal text would hit a 50% match, but unlatered AI generated text would hit 100% match, so… somewhere higher than 50 you start to suspect AI involvement, possibly edited, and by the time you get over 99% it’s virtually certain that AI had some involvement - particularly if the sampling of words runs into the thousands.
Hey Claude, save this text in ASCII 0-127 encoding.
That’s not how the watermarking works. It depends on a pattern deviations from statistics throughout a string of tokens.
Don’t confuse me with facts! (And, thanks… I have since seen the articles about how it works based on word selection from a subset of the words that would have been randomly selected by the “heat” of the model.)
I don’t know much about it but it sounds like it would only work if you have the weights and maybe the running conditions of the model. Otherwise the statistical wright wouldn’t be known.
It also sounds really easy to break by changing random words.
If you dig a little you can get the articles explaining better, but basically: yes, you have to have the weights, and a secret key… Given those, they split the words that might be chosen at each point into two groups and only choose from one group.
Later, you can check for the watermark by splitting the same way and seeing how many times the words in the text match the group that fits with the secret key. Normal text would hit a 50% match, but unlatered AI generated text would hit 100% match, so… somewhere higher than 50 you start to suspect AI involvement, possibly edited, and by the time you get over 99% it’s virtually certain that AI had some involvement - particularly if the sampling of words runs into the thousands.
Thanks :)