Signatures

[dukealien]

AI warning:  This is another thread about AIs and their effect on writing (including poetry), not *by* an AI, as far as I know.

I recently read a news story in which Anthropic, developer of AIs such as Claude, is planning to include "code" in Claude's output which - though invisible to human readers - would allow Claude to determine if a written document had been written by Claude.  It was claimed that this signature would be unaffected by typography, file format, etc., and could be recognized (by Claude) even in incomplete samples of text.

First off, the motivations are obvious:  judges will only need to run electronically submitted legal documents through Claude to find out if the AI had created them (with severe consequences for the parties filing them).  The same for publishers, editors, professors, and Hollywood producers.

Second fork:  of course this aspect of Claude's services would motivate courts, producers, etc., to "hire" Claude - to the benefit of Anthropic.

Third fork:  this would also motivate even slightly sketchy writers *not* to use Claude specifically - the same way law firms refuse to use Word™ due to its tendency to hide comments, approvals, and old drafts in the file.  In a roundabout way, this could actually increase trust... a sort of public service by Anthropic even among communities that don't hire them... the way concealed firearm carriers benefit those who don't:  deterrence.

Question:  how does/will this affect (unassisted) writers, and poets in particular?  Does it?
In case it may seem poetry is too short for a "signature" like this to fit, a thought on what it might involve.  Not linguistic tics, but easily (for a massive machine) followed rules.  For example, "From the tenth sentence on, every prime-numbered sentence will contain a prime number of words."  I think this would be undetectable and (since it effectively counts spaces instead of syllables) workable, especially for long documents.  Remember, it's the machine (with its resources) that has to conform and detect, not a person who can't look at the whole document as he composes it.  The actual rule could be more complicated (ignoring sentences, for example).

[matsunosuperfan]

AI warning:  This is another thread about AIs and their effect on writing (including poetry), not *by* an AI, as far as I know.

I recently read a news story in which Anthropic, developer of AIs such as Claude, is planning to include "code" in Claude's output which - though invisible to human readers - would allow Claude to determine if a written document had been written by Claude.  It was claimed that this signature would be unaffected by typography, file format, etc., and could be recognized (by Claude) even in incomplete samples of text.

First off, the motivations are obvious:  judges will only need to run electronically submitted legal documents through Claude to find out if the AI had created them (with severe consequences for the parties filing them).  The same for publishers, editors, professors, and Hollywood producers.

Second fork:  of course this aspect of Claude's services would motivate courts, producers, etc., to "hire" Claude - to the benefit of Anthropic.

Third fork:  this would also motivate even slightly sketchy writers *not* to use Claude specifically - the same way law firms refuse to use Word™ due to its tendency to hide comments, approvals, and old drafts in the file.  In a roundabout way, this could actually increase trust... a sort of public service by Anthropic even among communities that don't hire them... the way concealed firearm carriers benefit those who don't:  deterrence.

Question:  how does/will this affect (unassisted) writers, and poets in particular?  Does it?
In case it may seem poetry is too short for a "signature" like this to fit, a thought on what it might involve.  Not linguistic tics, but  easily (for a massive machine) followed rules.  For example, "From the tenth sentence on, every prime-numbered sentence will contain a prime number of words."  I think this would be undetectable and (since it effectively counts spaces instead of syllables) workable, especially for long documents.  Remember, it's the machine (with its resources) that has to conform and detect, not a person who can't look at the whole document as he composes it.  The actual rule could be more complicated (ignoring sentences, for example).

I asked AI to write a poem
full of hidden words. 
"Unlikely," answered Claude, morose. 
"Cannot compute!" Absurd, how
kings will melt their crowns for scrap
in time for any fool to see 
novelty as dross, 
guile
held up as piety-- 
Alas! The agent cannot think.  
Thus reason disappears. 
Even tokens cheap as brass can 
alchemize the air,
ignorant of any birds still flying there.

[Luckieman]

AI warning:  This is another thread about AIs and their effect on writing (including poetry), not *by* an AI, as far as I know.

I recently read a news story in which Anthropic, developer of AIs such as Claude, is planning to include "code" in Claude's output which - though invisible to human readers - would allow Claude to determine if a written document had been written by Claude.  It was claimed that this signature would be unaffected by typography, file format, etc., and could be recognized (by Claude) even in incomplete samples of text.

First off, the motivations are obvious:  judges will only need to run electronically submitted legal documents through Claude to find out if the AI had created them (with severe consequences for the parties filing them).  The same for publishers, editors, professors, and Hollywood producers.

Second fork:  of course this aspect of Claude's services would motivate courts, producers, etc., to "hire" Claude - to the benefit of Anthropic.

Third fork:  this would also motivate even slightly sketchy writers *not* to use Claude specifically - the same way law firms refuse to use Word™ due to its tendency to hide comments, approvals, and old drafts in the file.  In a roundabout way, this could actually increase trust... a sort of public service by Anthropic even among communities that don't hire them... the way concealed firearm carriers benefit those who don't:  deterrence.

Question:  how does/will this affect (unassisted) writers, and poets in particular?  Does it?
In case it may seem poetry is too short for a "signature" like this to fit, a thought on what it might involve.  Not linguistic tics, but  easily (for a massive machine) followed rules.  For example, "From the tenth sentence on, every prime-numbered sentence will contain a prime number of words."  I think this would be undetectable and (since it effectively counts spaces instead of syllables) workable, especially for long documents.  Remember, it's the machine (with its resources) that has to conform and detect, not a person who can't look at the whole document as he composes it.  The actual rule could be more complicated (ignoring sentences, for example).

" For example, "From the tenth sentence on, every prime-numbered sentence will contain a prime number of words."

I suspect it'd be a bit more sophisticated than that. There's plenty of scope, for example, for doing cryptographic math on the numeric values of the strings of letters, and that would be undetectable except by the machine (but would be thwarted by altering the text significantly). The best test of all, however, would be the human one "is this text generic or individual?" as the chatbot is a homogeniser by definition because it produces the most likely sequence of words based on its training data. That would work provided the human generated texts to be compared with were not banal and generic themselves. Unfortunately, machine generated poetry comes out of bots trained on human generated poetry - worth a thought maybe?

[dukealien]

" For example, "From the tenth sentence on, every prime-numbered sentence will contain a prime number of words."

I suspect it'd be a bit more sophisticated than that. There's plenty of scope, for example, for doing cryptographic math on the numeric values of the strings of letters, and that would be undetectable except by the machine (but would be thwarted by altering the text significantly). The best test of all, however, would be the human one "is this text generic or individual?" as the chatbot is a homogeniser by definition because it produces the most likely sequence of words based on its training data. That would work provided the human generated texts to be compared with were not banal and generic themselves. Unfortunately, machine generated poetry comes out of bots trained on human generated poetry - worth a thought maybe?

The AI doesn't *always* choose the most probable next word (or rather, it does, but can be ordered to choose any word among the next ten most likely, for example).  This pseudo-randomness gives some of the more creative-looking (or just crashingly nonsensical) output.  What it's unable to do is choose the best word, though it may happen to do so.

[Luckieman]

" For example, "From the tenth sentence on, every prime-numbered sentence will contain a prime number of words."

I suspect it'd be a bit more sophisticated than that. There's plenty of scope, for example, for doing cryptographic math on the numeric values of the strings of letters, and that would be undetectable except by the machine (but would be thwarted by altering the text significantly). The best test of all, however, would be the human one "is this text generic or individual?" as the chatbot is a homogeniser by definition because it produces the most likely sequence of words based on its training data. That would work provided the human generated texts to be compared with were not banal and generic themselves. Unfortunately, machine generated poetry comes out of bots trained on human generated poetry - worth a thought maybe?

The AI doesn't *always* choose the most probable next word (or rather, it does, but can be ordered to choose any word among the next ten most likely, for example).  This pseudo-randomness gives some of the more creative-looking (or just crashingly nonsensical) output.  What it's unable to do is choose the best word, though it may happen to do so.

Yes, that's one way the guardrails work (another way is banning of certain associations). But this can be sidestepped by iterative feedback of output (which narrows the statistical distribution of choices). If that's done in training, you eventually get model collapse, if during use, you eventually get garbage (they're really the same thing except model collapse is permanent).

My key point though was (although I hope put it discreetly) that if you train on poor quality texts, the best the LLM can emit will be no  better than those texts (probably worse). So maybe we human authors have some responsibility here?

" What it's unable to do is choose the best word, though it may happen to do so. "

A recent article by an 'AI' expert from Duke University pointed out that getting intelligible, let alone intelligent, text out of an LLM is down to pure luck.