Humanist Discussion Group, Vol. 38, No. 271.
Department of Digital Humanities, University of Cologne
Hosted by DH-Cologne
www.dhhumanist.org
Submit to: humanist@dhhumanist.org
Date: 2024-12-09 10:43:54+00:00
From: Tim Smithers <tim.smithers@cantab.net>
Subject: Re: [Humanist] 38.265: AI, poetry and readers
Dear Jim,
Thank you, again, for engaging thoughtfully and constructively
with my wonderings and wanderings. Your clarity helps, I
think, to talk about "generative" poems.
First. I must admit, I do feel out on a plank here, and one
extended from a ship I'm hardly qualified to be a passenger
on, let alone knowledgable enough to crew.
I'll start walking along this plank with something you say
which I agree with, mostly, from later on in your message.
"A poem is nothing other than a bunch of words arranged
into lines. If it looks like a poem, it is a poem. It
doesn't matter how it was produced. ..."
Yes, I agree, a poem is "a bunch of words," but not a bunch of
text. For me this difference is crucial. And, it matters
what, or better said, who does the looking here, and it
matters that there is a looker doing the looking, and one that
can see that they are looking at marks on a page that stand
for words. This is not what today's automated text generators
do; these just generate text. Not words. I agree that it
doesn't matter how the words of a poem are produced, but, to
me, it must be words that were produced, not just text. So
any text produce by an automatic text generator that looks
like the text of words that form a poem which has been
written, is not a poem; it is, I would say, a de-generative
poem.
I'll try to explain why and how, in what is a [very] long
post!
The first difference I see, from the end of the plank, shows
up here. You say ...
"I would never argue with Tim that some kind of human
intent is always behind the production of text. ..."
But I would not say this. I don't want to say this. My
difficulty is with the word "text." It should be, and, I
think, it needs to be, "words." I'd agree with saying "...
some kind of human intent is always behind the production of
words." All this so called Generative AI stuff is, I think,
clearly showing us that no intent, human or otherwise, is
needed to generate text, and text that is readable into words
by us. But, as you know, Jim, this is not the first time
people have made text generating systems to make words from.
A good example would be David Bowie’s and Ty Roberts’
Verbasizer program of 1995, I think: try it here
<https://verbasizer.com/>. What's different now is just a
matter scale. And, of course, that Bowie intended to use
[some of] the generated text as their own words in a creative
way, and was, I think, particularly good at doing this. Most
people who use automated text generators today don't do this.
They just read the generated text as if it is the result of
some words made with some intent, and either don't care that
there was no real intent, or presume that the Generative AI
systems they use actually have intentional capacities. Which
they don't!
Next, I'll dive in here, with your claim:
"So the meaning of the poem resides in the words on the
page, not in the poet who produced them. As a result,
there's no reason why an AI generated poem can't be
subject to the same kind of analysis as a humanly written
poem."
I agree with your first words, "So the meaning of the poem
resides in the words ...," and, I agree with your last words
of this first sentence, "... not in the poet who produced
them." But, I don't agree with what you put between these
words: "on the page." I'll get back to your second sentence
here after dealing with this first difference.
The words of the poem are not, I want to say, on the page.
They are in the heads of the human readers reading the text of
the poem which _is_ on the page, and they were, at some time,
in the head of the poet who formed them and then wrote them
down, as text on a page. I know this is an unusual
distinction, and one we don't usually make. But it is, I
think, a distinction we need when it comes to understanding
today's so called Generative AI systems, and, in particular,
automated text generation machines.
Words don't exist on the page, nor in the compression waves of
the air that allow us to hear the words we or others speak. I
want to say words, and thus whatever the words may mean to
people, only exist in the heads of human writers and readers,
and speakers and listeners. [Words perhaps also exist in the
heads of some other animals, but I'll leave this question
behind here.] Speaking and listening, and writing and
reading, and signing and seeing, and touching and feeling, (as
in braille), are thus ways we have of attempting to have the
words existing in one [human] head, also having an existence
in another [human] head, or lots of other heads. I call this
languaging. In the richness we experience this languaging, it
is a unique human quality. And it needs, I think, humans for
it to truly happen, so far, at least.
Seen this way, I think, means that to have any words you need
something with a capacity to form words and combinations of
words, to think with, and to say things with, perhaps just to
itself, but more usually to other things that also have word
forming and word using capacities, together with sufficient
hearing, seeing, feeling, and thereby reading and
understanding capacities. And, as you've probably
anticipated, I don't think Generative AI systems like Claude
3.5 Sonnet and ChatGPT have any capacity to form words which
they then write out as text for us to then read, so as to have
in our heads the words they first formed in order to say
something they intended to say. These generative systems
don't have any capacities for hearing, seeing, feeling,
reading, and understanding words, nor any intentions. They
only generate the marks of text, which, yes, we, with all our
languaging capacities, can read and understand things from.
But, no words have existed anywhere in the text generation
processes used by these systems. Text generating Generative
AI systems work only with very large numerical vector
representations of text-tokens [marks that make text] and
pre-built estimated probabilities for a great many [very]
detailed statistical text-token patterns found in the texts
left over from enormous amounts of human writing.
If we treat the text generated by things like Claude 3.5
Sonnet and ChatGPT as words, then, it seems to me, we are
saying, at least tacitly (!), that these systems _do_ have
word forming and using capacities, and thus that these
systems, like us humans, traffic in meanings, using all the
different languaging capacities this entails. And, some
people, I think, will want to say they do have word forming
and word using capacities like we humans have. The only way
of defending a position that says these machines do not have
anything like human languaging intelligence, and the
capacities that give us this intelligence, is to hold that
there is a distinction between words and text, albeit a
distinction we don't usually use, nor needed much before.
Here I'll return to the second sentence of your words I
started with above:
"As a result, there's no reason why an AI generated poem
can't be subject to the same kind of analysis as a
humanly written poem."
First, I want to insist that whatever it might look like to us
languaging beings, what we get from these AI text generators
is not what I am prepared to call a poem, not a real one, at
least. It's a plastic poem, as in plastic flowers. It may
look like the text of a poem but it isn't the text of a real
one, it's just something "made of plastic" to look like a poem
... because the system that generated this looks-like-a-poem
text has plenty of statistics about what the text of written
down poems look like as marks on a page, but not any notion of
what words are, what meaning is, nor of what a poem, as we
humans understand poems to be, nor any knowledge or
understanding of what languaging is. These Generative AI
systems have no real languaging capacities, just impressively
large matrix arithmetic capacities for crunching loads of very
large numerical vector representations of text-tokens and the
estimated probabilities of many detailed relationships between
text tokens.
But, yes, we language capacitated beings can indeed do the
same kind of analysis on the text generated by these machines
as we do on the text of humanly written poems -- that is,
those of us qualified to do this can. But, when we do this,
we first need to construct words from the text we read, and,
from these words, form meanings, which are needed in our
analysis efforts. But, if we do this, we should, I think,
acknowledge that the text we read, and form words from, and
thus meanings from, was not the result of any writing of words
by a poet: no words pre-existed the text we read. So, all the
word and meaning forming we do in our analysis are the first
words, and meanings, formed. This means, I would say, that
this kind of analysis is different from the analysis of a real
poem written with intent by a poet. In this analysis of some
text as if it is text from the forming and writing of a poem,
necessarily involves, perhaps tacitly, the forming of a poem
to analyse. It is you who do this analysis of automatically
generated text, that makes the poem you analyse, and you have
to do this to have any poem to analyse; to have an words, and
thus meanings, to work with. Real poems formed and written by
a poet come with words. Automatically generated text does not
come with words, only text. You have to supply the words,
which makes you the poet here, I would say.
I think this difference matters because the machine that
generated the text we take to be the text of the written words
of a poem, has none of the capacities needed to do even a poor
version of this analysis. But that won't stop it generating
text that looks like the outcome of some analysis, if we
prompt it to generate text in this way. These Generative AI
systems have a seemingly unlimited capacity for a kind of
artificial pretending. They will generate no end of text that
looks a lot like text that resulted from someone writing
words. But no words were ever made.
So, suppose we do such an analysis of the text generated by
some machine, and thus treat this generate text as if it is
the text of a poem that some one formed and wrote down. What
will we get? Well, the analysis of a poem, of course. But it
is only the analysis of a poem because we, who did the
analysis, made the generated text into a poem, not because the
text we chose to analyse is the text of a poem that was made
and written down by someone. There was no poem making and
writing down in the generating of the text we analysed. We
just chose to treat the text as if there was. Any poem that
results is made by us who do this analysis. There is no poet
here. This is, I would say, analogous to someone doing a
botanical analysis of an artificial flower, just as they might
do when studying a real flower. Nothing stops the botanist
from doing this, except, perhaps, honesty, but no amount of
botanical analysis will turn the artificial flower into a real
flower; there wasn't a flower there to start with. The same
goes for the automatically generated text that, to us, looks
like the text of a written down poem; no amount of analysis of
this text will turn it into a poem a poet wrote, no matter
what this analysis might uncover in the words we form from a
reading of the text, words, I repeat, which were not there
until we did the reading, word forming, and analysis.
What we get, and can only get, from automatic text generators
is text that has never been words in anybody's head. And
that, I think, makes a big difference, an important
difference, a difference we should always acknowledge, and not
ignore, or forget, of seek to hide, or pretend doesn't matter.
Not unless, that is, we can confidently show how all that goes
on in our heads when we language is something rather like
matrix arithmetic using very large numerical vectors shown to
reliably represent, and only represent, text-tokens, and the
estimated probabilities of how they relate, and, of course,
that we can also explain how all this statistical matrix
arithmetic came to be in our heads in the first place. [All
of which just happens to be something the Connectionists
don't' appear to feel a need to explain. But why would they
when you just need to accept, and then run with, the
Connectionist dogma: build lots of connected things you call
"neurones" then it must be intelligent, and intelligent like
us.]
Before stopping I want to say more about this "strange
distinction" I make here: it's text we have on the page, not
words; words only exist in the heads of the author and the
reader(s): the author who wrote the words of their poem, and
the reader who forms the same words from their reading of the
text, but who does not necessarily have the same meanings from
these words as the author had when they wrote them, or, even
have any meaning at all, perhaps.
This distinction is not really all that strange, I'd say.
Take a musical score, for example, for the last movement of
Beethoven's 7th symphony, say. We don't usually say the score
is the music, that the music is on the page. It's marks that
are on a page, not the music. If it was the music on the
page, we'd just need to give people printed copies of this
score and they'd have the music. But they don't. Performance
is needed; performance by suitably qualified, skilled, and
practiced, musicians. And, just like text can be read into
words that mean different things to the different people who
read the same text, so a musical score can be, and just about
always is, performed differently, and this difference matters.
These different performances, interpretations as we call them,
are taken to be a kind of exploration of what the written
score can be used to make. And, mostly, we take it that the
composer who first wrote the score intended it to be used in
this way, albeit with some indications and occasionally
further instructions, on how they want their work to be
performed, but not interpreted: composer can't instruct this,
and don't try to, of course.
Similarly, a Laban notation score -- using the notion invented
by Rudolf von Laban to record and thereby analyse human
movements and motions, and, in particular, dance movements and
motions -- is not the dance, it's marks on a page. Once
again, performance is needed, and interpretation is needed, to
have the real [notated] dance.
Semiotics. This is all semiotics. Which depends upon a clear
distinction between the sign and the signed or signified. The
art of semiotics lies in accurately identifying the signified
from the sign, and thus what a particular sign is intended to
mean, or can be reasonably understood to mean, and, more
interestingly perhaps, what some combination of signs may be
reasonably taken to mean. This art includes the need to
identify when we have a empty sign, something we have been
told, perhaps, or believe, to be a sign, but which isn't a
sign; it doesn't signify anything, but just looks like it
does. Empty signs are what automatic text generators
generate, but they make these empty signs to look so like
content-full signs that we nearly automatically read them as
such, and make words from our reading of these signs, so that
the combinations of the words we form say things to us, but
all these words and what they say to us are in our heads, and
only in our heads, and are all made by us; they do not come
from the page. Only text is on the page, only signs, and, in
this case, only empty text is there. It takes us, and our
languaging capacities and skills and imagination to make words
in our head from this empty text, or what we might call
word-less text. And word-less text results in de-generative
poems, I think. Not real poems.
I've used the phrase "in the head" lots here, to say words
only exist in the heads of people. This is, of course, a
short-hand for all the cognitive goings on when we humans
language, and about which I don't want to say much here. We
know some about all this, but not everything, and, perhaps not
as much as we like to think we know. These cognitive goings
on are the proper realm of a wide variety of kinds of
researchers, in the Cognitive Sciences, Neurosciences,
Psychology, Anthropology, Linguistics, Philosophy of Mind and
Language, including Epistemology, AI [the research discipline]
and, of course, Literature and Literary Studies in all its
forms, to name the most obvious ones, but probably not all.
I've treated all this as being wrapped up in a Blackbox I call
the head of a person.
Still, there is one thing I want to say something about, of
what goes on inside these Blackboxes; about the nature of
words as they exist in our heads. They are not simple things;
they are not mere simple symbols, or signs. They are, it
seems to me, and must be, complex things; things which have
important properties and qualities of their own, and which
have internal state and structure, in order to work as what we
call words, to think with and to say things with, and make
poems from, as we do.
For one thing, words, and thus what we take them to mean, or
use them to mean, often have a fogginess, or mistiness, that
text does not, and must not, have. Text, to work properly,
needs to be reliably readable; we need to be able to
accurately read the text so we can re-construct the words that
were written and which therefore left the marks on the page we
read them from. But this fogginess or mistiness is important;
we need what we want our words to mean or what we take the
words we reconstruct from text to mean, to be a little
undefined, mouldable, flexible, slippery, suggestive, ... It
is, I think, this lack of definitiveness of words that makes
languaging possible. If we had to be definitive in every word
we used, and in every combination of words we form and use,
languaging would fail so often we wouldn't use it much; errors
in speaking or writing the definitive words, and errors in
(re)constructing the definitive words from what we hear or
read, and thus errors in the definitive meanings of the words,
would occur too often for our word based interactions, ie,
languaging, to be viable.
This is why programming computers, which looks rather like a
kind of languaging, and which we often talk of this way, is so
different from human natural languaging. The "words" and
constructions of programming languages, and programs, do need
to have definitive denotations for us to be able to use them
to build programs computers can execute. And, as a result, we
often make mistakes in the forming of these "words" and
correct combinations of them. We call these mistakes "bugs,"
a word which has lots of useful slipperiness and flexibility,
as demonstrated in this use here!
How is languaging between us viable is also a question I don't
want to get into here; it's another long story, I think. But,
the way I see it, and have approached it -- in designing and
implementing an artificial language for people to use to
interact with a robot; a "spoken" language built from sounded
musical notes which the robot used to say [simple and limited]
things to people, and which it could hear and understand
itself say, and which people could [learn to] use to say
[simple, and limited] things to the robot -- is that
languaging is necessarily a situated activity which requires
enough shared common situation and conditions, which therefore
makes it better treated as a distributed cognition phenomenon,
and not just a "in side the head" phenomenon, as I have
indicated here with my Blackbox treatment.
The nature and structure of words as they exist in our heads,
and as we use them in our thinking and languaging, are, I
think, less studied than they need to be for a proper
understanding of how languaging works. Words, in our heads,
are not just simple symbols for meanings which we select and
combine to capture and express what we want to say and to mean
-- though this is an impression we may get from looking at
what linguists say. I don't see it as being nearly as simple
as this. And, I think this from the work I, and many others
in Symbol Processing AI, have done on building computational
systems that do know, understand, and reason about certain
things. Yes, this does involve processing symbols, but these
symbols often have to be more than simply tokens that signify
well defined, and fixed, things. The symbols that get
processed by the computational mechanisms we have to design
and implement to get any actual knowing, understanding, and
reasoning to happen, have to have complex internal modifiable
structures and variables, which are, themselves processed,
manipulated, and changed, by the computations of the the
knowing, understanding, and reasoning, we build our systems to
do. This is a reality of so called Symbol Processing AI often
missed, including by people who do this kind of AI work, but
who don't reflect much on what they have built and why it does
what it does, if, indeed, it actually does what they say it
does. John Searle, and The Chinese Room Argument they
presented in 1980, suffers from this failure to understand the
real nature of symbols in AI systems, and that they are not,
and can't just be, simple signs for things to have some
intelligent behaviour.
But, notice, John Searle was right about the processing of
[simple] symbols not being enough for real understanding, but
this is all we have in automated text generation systems;
loads of processing of numerically represented text-tokens, a
kind of large but nonetheless simple symbol, but often dressed
up to look like it has lots of clothes on. There are no
clothes, I think.
If we accept automatically generated text to be the same as
the intentional forming and writing of words we will let these
so called AI systems win, and win without even being
intelligent. It is, as always, up to us humans to avoid
dehumanising ourselves.
There's more to say, but I'll stop here, and thank you, Jim,
and anybody else who has made it down to here, for reading all
this text, and for all your efforts of making words in your
heads from it all. I'll be interested to see if the words you
make from this text mean anything to any of you all?
-- Tim
PS: Willard. I hope I've not broken the record for the most
amount of text ever posted to the Humanist list?
> On 2 Dec 2024, at 08:36, Humanist <humanist@dhhumanist.org> wrote:
>
>
> Humanist Discussion Group, Vol. 38, No. 265.
> Department of Digital Humanities, University of Cologne
> Hosted by DH-Cologne
> www.dhhumanist.org
> Submit to: humanist@dhhumanist.org
>
>
>
>
> Date: 2024-12-01 19:46:47+00:00
> From: James Rovira <jamesrovira@gmail.com>
> Subject: Re: [Humanist] 38.263: AI, poetry and readers
>
> Once again, Tim prompts me to respond. My own background includes work in
> poetry: half of my dissertation was about a poet, William Blake, and I've
> taught both formal and free verse at the undergraduate and graduate levels,
> and I've supervised creative theses that were collections of poetry,
> besides having published a number of poems myself. That's my background.
>
> I would never argue with Tim that some kind of human intent is always
> behind the production of text. Most literary scholarship disregards intent
> as inaccessible or irrelevant, however. The fact that a poem, even one
> generated by a prompt fed to a computer, is an intentionally made product
> usually doesn't help us understand it. If we do enough biographical and
> historical research, and if we even know the author and the time and
> conditions of the writing of the poem, that will help us understand the
> author at the moment of writing the poem, but not what the poem in any
> absolute sense *means.* T.S. Eliot's essay "Tradition and the Individual
> Talent" is a good reflection on the relationship of poem to author from the
> author's point of view:
> <https://www.poetryfoundation.org/articles/69400/tradition-and-the-individual-
> talent>.
>
> So the meaning of the poem resides in the words on the page, not in the
> poet who produced them. As a result, there's no reason why an AI generated
> poem can't be subject to the same kind of analysis as a humanly written
> poem.
>
> Before I move on, I'd like to define a couple of terms: Formal verse
> follows predetermined rhyme and metrical patterns (at least one or the
> other, often both) while free verse does not. Free verse may follow such
> patterns, but they are invented by the poet for the poem.
>
> What I'm going to say is about formal verse first: I love the invocation of
> Lewis Carroll at the end of Tim's post, who was a mathematician. I believe
> that background in mathematics contributed to the excellence of his verse.
> I frequently compare a line of poetry to a bar of music (and I'm hardly
> original in doing so), and the mathematical qualities of a bar of music is
> hardly contested. Neither should the mathematical qualities of a line of
> formal verse. Many kinds of formal verse forms consist of a fixed pattern
> of stressed and unstressed syllables often combined with a fixed rhyme
> scheme.
>
> As a result, the tendency, or danger, is for the poetic form to turn the
> poet into a computer rather than for the poet to own the form as meaningful
> self-expression. As a result, there's no reason that a computer couldn't
> produce something that at least fits the pattern, even a conceptual pattern
> that requires a certain kind of rhetorical gesture around, say, line 9 of a
> sonnet, or the last two lines of an Elizabethan sonnet. As a result,
> there's no reason why the computer generated poem couldn't be subject to
> analysis *as if* a human being had written it. It is still made up of human
> words.
>
> A poem written by a computer as an intentionally devised means of authentic
> self expression would probably be in binary, or in electrical impulses, or
> some kind of representation of them. If computers really wrote their own
> poetry, only a programmer could read it, or an engineer with an amprobe and
> a volt meter.
>
> Just to be clear, I do not believe this will ever happen.
>
> HOWEVER, once you begin writing *parodies*, you turn even a free verse
> source poem *into a formal verse form*. It suddenly has fixed parameters,
> even for specific conceptual moves and rhetorical devices. And above all
> else, a good parody is an inherently bad poem imitating a good one, and the
> better it is at imitation *while still being a bad poem*, the better a
> parody it is. So I think AI is especially suited for writing parody poems.
> I only find it sad that it can't laugh along with us.
>
> What we're circling around without discussing, however, is what a poem
> fundamentally *is*.
>
> I think that many people believe poetry is a deeply meaningful and symbolic
> means of human self-expression.
>
> Every bad human poet who has ever lived thought just that. Douglas Adams
> thought just that in a very funny way. Poets who focus on their own
> feelings and ascribe inherent worth to them while they are writing their
> poems massively suck, both as human beings and as poets.
>
> Whenever poetry rises to meet that criteria, it is only as a by-product.
>
> A poem is nothing other than a bunch of words arranged into lines. If it
> looks like a poem, it is a poem. It doesn't matter how it was produced.
> That is the hard and cruel truth. Eliot faced that truth; Milton faced it.
> The poem doesn't care about the poet. The poem does not express the poet.
> The poem expresses itself. The poem *leaves* the poet and then does
> whatever it will, and if it's a really good poem, it means a lot more than
> the poet ever hoped to express. So good poets really only care about the
> words on the page. "Care" is an emotion, and there is probably another
> emotion behind that care, but the focus of their attention is on the artful
> arrangement of words on a page or screen. If the focus is anywhere else,
> the poem is bad.
>
> A poem is not soft and meaningful and of deep feeling. A poem is hard and
> cruel and meets you only on its own terms. It does not know you or care
> about you. It does not feel: it may provoke feelings, but it *doesn't care
> about your feelings*. It simply is. If a poem became sentient, it'd be a
> real jerk. So good poets very carefully *construct*, but not *write*,
> poems. A poem is a kind of emotional/conceptual program that, once arranged
> by a human poet, works outside of its own parameters.
>
> So I disagree with this claim: "But, looking-like a poem is not a
> sufficient, nor even a necessary, condition of being a poem."
>
> I think that is the only condition of being a poem. I think I would agree
> with Tim's claim if we were describing a *great* poem. I don't think AI can
> do that, only because it would need to understand nuance incorporating too
> much outside the poem, beyond syntactical patterns.
>
> But most human beings who write poetry actually write bad poetry. Some
> write decent poetry, even good poetry. Most will never write great poetry
> ever, at all, their entire lives. But if they're lucky, they might write a
> few great lines here and there. That makes all the rest of it worth it.
>
> Jim R
>
> On Sun, Dec 1, 2024 at 3:01 AM Humanist <humanist@dhhumanist.org> wrote:
>
>>
>> To be clear. It takes a person to decide to present some text
>> as a poem, whether they wrote the words, and are thus left
>> with the text marks of their words, or got the text generated
>> by some gee whiz machine. I bet if the generated text didn't
>> look somehow like a "poem" to this person, they wouldn't
>> present the text as a "poem." But, looking-like a poem is not
>> a sufficient, nor even a necessary, condition of being a poem.
>> Trying to make "poems" using text generators thus easily
>> results in what I call 'Artificial Flower' "poems" -- texts
>> which may look like poems to many people, but ain't poems, not
>> real ones -- and, cannot result in things I would call
>> 'Artificial Light' poems -- marks that are artificially made
>> somehow to be used to recreate words well chosen to form real
>> poems, and which may not look like any poems we've seen
>> before, but, which, upon genuine analysis and interpretation,
>> can be shown to be, and thus appreciated as, real poems,
>> perhaps in new and exciting ways.
>>
>> -- Tim
_______________________________________________
Unsubscribe at: http://dhhumanist.org/Restricted
List posts to: humanist@dhhumanist.org
List info and archives at at: http://dhhumanist.org
Listmember interface at: http://dhhumanist.org/Restricted/
Subscribe at: http://dhhumanist.org/membership_form.php
