Maybe I’m out of the loop, but how would I watermark a text without sounding obvious?
By using weird phrases? They get edited. By using e.g. an exact combination of starting letters over a large paragraph? One changed word and it’s broken. And even if not, I could happen to write the same myself, and then?
How am I hiding a signature in Plain text? Anyone got a better idea than my silly ones?
The more apt word is steganography, rather than watermark. Basically subtly adjust the weights of the model so that some subtle patterns appear. Think of how AI text prefers certain words and phrases that ordinary humans don’t use as often, like “delve,” but presumably much more subtle.
And no, as Anthropic has already said, this watermark may not survive editing/formatting.
By being just a little weird, but in a pattern, and not to you. A pattern like: every n tokens raise the temperature (randomness) for one token. Then to dectect AI, you tokenize and calculate how expected each next token is. Then you try fitting the pattern to that.
To my surprise almost if not all LLM are set not to be deterministic and have one unique input result in always the same output.
They all are set to have a “temperature” setting so that they sound more natural.
Personally I think quality of the output tokens of LLM is surprisingly not as much their priority as the quality and truthfulness of the result.
I would 100% prefer a LLM that is purely deterministic and repeats the same answer exactly to the same question. Instead LLM are constantly choosing the next likely token more TL appear human rather than being accurate.
These LLM are designed as sycophants and set up and trained as such.
So a fingerprinting in the output seems quite realistic. An LLM is not giving you it’s best most likely answer. It’s taking one of the most likely answer and adds a sprinkle of uncertainty and randomness on top of it so it looks natural…
Sounds fitting, but we’re talking of text with a purpose here. Wouldn’t we notice that raise in temp?
If I “write” a novel like that I will surely still cross-check it a dozen times, no?
Also I might rephrase a lot, move a lot, and the confidence of the detection would fall.
But, yes, on a long text that I would not touch, that could work pretty reliable.
Lower temperature is not necessarily better quality. At 0 it will get very repetitive, it’s for classification tasks, not for prose.
I suppose the optimal temperature depends on the model and the task, and I don’t think it’s very sensitive. Varying temp might even give better results, who knows.
I used mullvad browser and mullvad VPN. Both of which I have never used previously. Through this I then created a proton mail account with a username I have never used. Signed up to Reddit with that. There is no link there.
It’s possible they banned me just because doing that in itself was just too shady but I’m not sure. Next I’ll try buying a karma’d account.
Probably just banned because vpn, especially proton. I can’t count the sited I can’t visit anymore because of VPN.
Even steam asks me questions like WHY am I signing in “like that”…
I’d try another VPN and/or just another fresh browser. One you’d never use anyway. Just to see where’s the culprit.
Maybe I’m out of the loop, but how would I watermark a text without sounding obvious? By using weird phrases? They get edited. By using e.g. an exact combination of starting letters over a large paragraph? One changed word and it’s broken. And even if not, I could happen to write the same myself, and then?
How am I hiding a signature in Plain text? Anyone got a better idea than my silly ones?
The more apt word is steganography, rather than watermark. Basically subtly adjust the weights of the model so that some subtle patterns appear. Think of how AI text prefers certain words and phrases that ordinary humans don’t use as often, like “delve,” but presumably much more subtle.
And no, as Anthropic has already said, this watermark may not survive editing/formatting.
Claude Now Watermarks Your Text | Vanja Petreski - https://vanja.io/claude-invisible-watermark/
By being just a little weird, but in a pattern, and not to you. A pattern like: every n tokens raise the temperature (randomness) for one token. Then to dectect AI, you tokenize and calculate how expected each next token is. Then you try fitting the pattern to that.
Wouldn’t that potentially affect output quality particularly in code/tool calls?
To my surprise almost if not all LLM are set not to be deterministic and have one unique input result in always the same output.
They all are set to have a “temperature” setting so that they sound more natural.
Personally I think quality of the output tokens of LLM is surprisingly not as much their priority as the quality and truthfulness of the result.
I would 100% prefer a LLM that is purely deterministic and repeats the same answer exactly to the same question. Instead LLM are constantly choosing the next likely token more TL appear human rather than being accurate.
These LLM are designed as sycophants and set up and trained as such.
So a fingerprinting in the output seems quite realistic. An LLM is not giving you it’s best most likely answer. It’s taking one of the most likely answer and adds a sprinkle of uncertainty and randomness on top of it so it looks natural…
Sounds fitting, but we’re talking of text with a purpose here. Wouldn’t we notice that raise in temp? If I “write” a novel like that I will surely still cross-check it a dozen times, no?
Also I might rephrase a lot, move a lot, and the confidence of the detection would fall.
But, yes, on a long text that I would not touch, that could work pretty reliable.
Lower temperature is not necessarily better quality. At 0 it will get very repetitive, it’s for classification tasks, not for prose.
I suppose the optimal temperature depends on the model and the task, and I don’t think it’s very sensitive. Varying temp might even give better results, who knows.
Probably the same way Reddit figures out I’m permabanned even if I remove all traces of my identify during signup.
You probably just only removed all traces of your identity that you can think of 😁 Though, nothing much lost. .
I used mullvad browser and mullvad VPN. Both of which I have never used previously. Through this I then created a proton mail account with a username I have never used. Signed up to Reddit with that. There is no link there.
It’s possible they banned me just because doing that in itself was just too shady but I’m not sure. Next I’ll try buying a karma’d account.
Probably just banned because vpn, especially proton. I can’t count the sited I can’t visit anymore because of VPN. Even steam asks me questions like WHY am I signing in “like that”…
I’d try another VPN and/or just another fresh browser. One you’d never use anyway. Just to see where’s the culprit.
Maybe I’ll try some kind of proxy. I have to use something because I’m IP banned for sure.