giusto oggi nel blog cultureIA (https://cultureia.hypotheses.org/) ho
visto la notizia di questo seminario tenuto il 28 gennaio di quest'anno:
/Artificial and Post-Artificial
Writing/(https://cultureia.hypotheses.org/3159)
disponibile in youtube (https://www.youtube.com/watch?v=wQKBxgiRIdA)
in cui la relatrice apre con questi pensieri:
"How do we read a text when we can no longer be sure that it was not
written by an AI?
With AI permeating writing tools and producing vast amounts of text, it
may then become impossible to distinguish artificial from natural text,
but also impossible to bear this constant uncertainty.
I speculate that we may thus eventually reach a new post-artificial
situation in which the standard expectation of text is replaced by an
agnostic stance in regard to its origins.
literary text which emphasizes a stronger notion of authorship may
resist this transition longer through deliberate linguistic
experimentation or by emphasizing a human maker.
By extrapolating from the current technological situation, the talk
attempts to think through some of the possible consequences of large
language models and related technologies for interpretive situations and
hermeneutic strategies at a time when we can no longer assume that any
reasonable, complex, and coherent text was written by a human."
la trascrizione che allego è prodotta con (mac)whisper.
Maurizio
Il 18/05/25 22:40, maurizio lana ha scritto:
ciao Daniela,
la questione in gioco - distinguere il testo prodotto da un chatbot da
quello prodotto da un umano - ne ha sullo sfondo una seconda -
individuare le caratteristiche che distinguono la scrittura di un
umano da quella di un altro (usualmente chiamate "stile"). la seconda
è al cuore degli studi attribuzione con metodi quantitativi che hanno
una lunga storia (il resoconto più ampio e valido è Grieve, Jack
William. «Quantitative authorship attribution: A history and an
evaluation of techniques». Thesis, Department of Linguistics - Simon
Fraser University, 2005. http://summit.sfu.ca/item/8840) e anche dei
successi significativi.
chi ha lavorato a studi di authorship attribution è portato in primo
momento a pensare che se è possibile individuare le caratteristiche
che distinguono la scrittura di un umano da quella di un altro, sia
possibile anche individuare quelle che che distinguono la scrittura di
un umano da quella di un chatbot.
ma in effetti il tuo commento mostra un'impossibilità in linea di
principio, perché nel chatbot non c'è una linea di pensiero ma un
mosaico di pezzi di altri. e dunque i deboli tentativi dei software si
collocano nell'individuazione di aspetti del tutto
esteriori/estrinseci sui quali un ipotetico falsario testuale può
benissimo e facilmente operare per modificarli/occultarli.
peraltro vedendo già adesso circolare grandi quantità di non-testi ,
ovviamente si pensa se ci sia un modo selezionarli e distinguerli
senza doverci spendere tempo leggendoli - che francamente si vorrebbe
poter spendere meglio il proprio tempo
e quindi mi pare che almeno come comprensibile anche se irrealizzabile
desiderio la questione rimanga viva.
quella per cui da un lato dobbiamo distinguere, dall'altro si vorrebbe
fare di melgio che leggere non-testi per poter dire che sono tali.
Maurizio
Il 18/05/25 00:24, Daniela Tafani ha scritto:
il senso mi pare quello di qualsiasi circonvenzione di incapace.
A questa collaborano studiosi di ogni sorta, a pagamento e non.
Si oppone, per ora, la FTC:
<https://www.ftc.gov/news-events/news/press-releases/2025/04/ftc-order-requires-workado-back-artificial-intelligence-detection-claims>
Per distinguere automaticamente il prodotto di un estrusore di stringhe di
testo probabili da un testo scritto da un essere umano, i due testi dovrebbero
avere qualcosa di diverso, che fosse rintracciabile da un software.
La differenza principale tra i due testi è che uno è stato pensato da chi l'ha scritto e
l'altro no, essendo un'immagine sfuocata di altri testi (quelli sì, pensati da qualcuno).
Suppongo perciò che serva il pensiero, per riconoscere un testo estruso da un software
che produca linguaggio senza pensiero rispetto a un testo prodotto da un essere pensante.
Non esistono, oggi, software pensanti, perciò gli "AI detector" non esistono.
Che ne pensi?
Per avere una differenza rintracciabile automaticamente, dovresti marcare in
qualche modo il testo sintetico, che però, immagino, potrà comunque essere
copiato come solo testo.
Grazie.
Un saluto,
Daniela
________________________________________
Da: nexa<[email protected]> per conto di Enrico
Nardelli<[email protected]>
Inviato: sabato 17 maggio 2025 19:53
A:[email protected]
Oggetto: [Junk released by Allowed List] [nexa] Qual è il senso...?
... di realizzare strumenti per la verifica che un testo sia stato generato
dall'Intelligenza Artificiale, visto che (oltre a sbagliare clamorosamente su un mio
testo scritto interamente da me che è stato attribuito per il 71% all'IA - ma l'ultima
volta che mi sono controllato mi son trovato sempre fatto della solita materia
organica... 😂) poi la stessa società offre uno strumento per "umanizzare" il
testo?
Io ho provato GPTZero, giusto per fare nome e cognome, ma penso che altre
startup simili mettano a disposizione prodotti simili.
Mi sa tanto di un gigantesco gioco delle 3 carte...
------------------------------------------------------------------------
many of us believe the EU remains
the most extraordinary, ambitious, liberal
political alliance in recorded history.
where it needs reform, where it needs to evolve,
we should be there to help turn that heavy wheel
Ian McEwan, The Guardian, 2/6/2017
------------------------------------------------------------------------
Maurizio Lana
Università del Piemonte Orientale
Dipartimento di Studi Umanistici
Piazza Roma 36 - 13100 Vercelli
------------------------------------------------------------------------
l’imbarazzante senso di colpa per la nostra vita
che va avanti coi suoi momenti di gioia e svago
mentre quella di altri non ne ha
S. Vizio
------------------------------------------------------------------------
Maurizio Lana
Università del Piemonte Orientale
Dipartimento di Studi Umanistici
Piazza Roma 36 - 13100 Vercelli
Hello everyone, hello, hello, I'm very happy to meet you virtually.
So this seminar is titled on artificial and post-artificial text, machine
learning and the reader's expectations of literary and non-literary writing.
And today's presenter is Annes Beja, an assistant professor of German at the
University of Berkeley of California.
And his research focuses on the history of German philosophy in the 20th
century, political theory and theories of the digital and AI.
So Beja's academic texts have appeared in configurations, poetics today and a
new German critic among others.
And his most recent work is digital literature.
So I am for wrong pronunciation.
I don't I don't speak German.
So with Simon Roloff published in Hamburg in 2024.
And Beja is also active as a writer of digital literature.
And also one of his most important works is the novel Berlin, Miami, published
in Berlin in 2023, which has which was co-written with the self-trained large
language model.
So we are very happy to to listen to your presentation.
Thank you again for participating.
And now the floor is yours.
Thank you.
Thank you so much.
Thanks for the invitation.
Thanks for the introduction and for the opportunity to speak.
So, yeah, my my presentation is called an artificial and post artificial text.
And I want to talk about the the the effects on the reception side of the
production of the synthetic production of text.
So it's been an eventful two years since Chats GPT was published in November
2022.
And although the technology underlying this large language model has been
around since 2017, Chats GPT was really the first time an LLM has been easily
accessible not only to the privileged few, but to the public at large.
Now, in other words, it was able for everyone to experience and to get used to
neural network based text synthesis to emulate human writing.
No matter how one gauges the perils or merits of large language models, it
seems likely that they herald in an era in which the text we encounter may be
entirely generated by a machine.
Practitioners of the hermeneutics disciplines then are called to consider what
impact the current rapid advances in machine learning research might have on
the way in which we understand and interpret text.
So in this talk, I want to contribute to an answer by rephrasing and thus maybe
delimiting the question a little bit.
What will be the impact of artificial writing on the readers' expectations of
unknown text?
I therefore turn to the standard reception of generative AI's output and I
shall answer two facets of this question.
First, what happens when we are confronted with artificial text alongside
natural ones, those who are written in the traditional way?
How, in other words, do we read a text when we can no longer be sure that it
was not written by an AI?
And second, what direction might this increasing doubt as to a text's origin
take if at some likely point the distinction between natural and artificial
itself becomes obsolete, so that we can no longer even seek to differentiate
and instead read post-artificial text?
I will discuss the impact of AI-generated text on the reader's expectations in
three steps.
First, I trace its origins back to early computerized text experiments of the
1950s and 60s, when the assumption that humans are the originators of text
becomes apparent because for the very first time an alternative to it exists.
I suggest that AI development has implicitly used this standard assumption of a
text's human origins by at least in principle aiming to deceive users into
thinking they are interacting with humans.
A second phase then is reached as AI text generation advances, producing
increasingly naturalistic outputs, and readers move from simply assuming a
human behind a text to actively doubting the text's origin.
With AI permeating writing tools and producing vast amounts of text, it may
then become impossible to distinguish artificial from natural text, but also
impossible to bear this constant uncertainty.
I speculate that we may thus eventually reach a new post-artificial situation
in which the standard expectation of text is replaced by an agnostic stance in
regard to its origins.
Such text, I argue, would tend to be read as authorless by default.
Finally, I discuss how literary text which emphasizes a stronger notion of
authorship may resist this transition longer through deliberate linguistic
experimentation or by emphasizing a human maker.
By extrapolating from the current technological situation, the talk attempts to
think through some of the possible consequences of large language models and
related technologies for interpretive situations and hermeneutic strategies at
a time when we can no longer assume that any reasonable, complex, and coherent
text was written by a human.
Now, that there is a standard expectation that readers have when confronted
with unknown text at all, namely that they are "natural", written by humans,
becomes observable only once there is an alternative to it, when they are also
artificial texts.
The distinction between natural and artificial text is not mine.
It was introduced by a German philosopher and author, Max Benze, in his 1962
essay on natural and artificial poetry.
While arguably too strict, probably no text is purely natural or artificial, it
provides a coherent articulation of the difference between human and
computer-written text, and it allows one to see where this difference starts to
collapse in the present.
In his essay, Benze considers how non-intentional, computer-generated
literature differs from intentional literature written by humans.
He focuses on the mode of creation behind these texts, what happens when an
author writes a poetic text.
For Benze, this is clear in the case of natural poetry.
For a text to have meaning, it must also be linked to the world via a personal
poetic consciousness, as he writes.
For Benze, language is largely determined by an ego relation and a world aspect.
Speech emanates from a person, no matter what they say, that person is always
the one speaking.
At the same time, in their speech, the speaker always refers to the world.
Poetic consciousness then puts being into science, that is, the world into
text, and ultimately guarantees that one is related to the other.
Without this consciousness, Benze holds, the science and the relationship
between them would be meaningless.
For a computer, words are only empty symbols, operative variables that are
devoid of intrinsic meaning.
It is precisely this case that Benze's second category, artificial poetry,
describes.
By this he means literary texts that are produced through the execution of a
rule, an algorithm.
In them, there is no longer any consciousness and no reference to an ego or to
the world.
If these texts mean anything, then only accidentally so and only for a human
reader.
Instead, such texts have purely material origins.
They can be described only in terms of mathematical properties such as word
frequency, distribution, degree of entropy, and so on.
The subject of an artificially generated text, then, even if the words should
happen to designate things in the world for us, is no longer actually the
world, but only that text itself.
It is the measurable, calculable, schematic object of an exact textual science.
If natural poetry, then, originates in the realm of understanding, artificial
poetry is a matter of mathematics.
It does not want to and cannot communicate, and it does not speak of a shared
human world.
Benze's thrust, however, is not to rescue a romantic idea of an inexplicable
human creative power by setting it off against Soler's computer.
On the contrary, the author as genius is dead and buried here.
Instead, Benze wants to know what can still be said aesthetically about a text
if one disregards traditional categories such as meaning, connotation, or
reference.
The answer he presents in his Information Aesthetics, which is strictly
positivist, and the tradition of Shannon and Weaver's communication theory,
considers only statistically measurable textual properties.
Artificial poetry, then, precisely because it is meaningless, is also pure
poetry.
It gets by entirely without the assumption of an underlying consciousness and
is an independent aesthetic object that can be investigated imminently.
Benze himself was involved in several experiments with artificial poetry.
The most famous of these was certainly the stochastic texts, which his student
Theo Lutz produced on the Zuse Z22 mainframe computer at the University of
Stuttgart in 1959, in which can be considered the first German-language
experiment with digital literature.
These texts are stochastic because they are randomly selected and assembled
from a given collection of vocab words.
The fact that these words are taken from Franz Kafka's "castle" hardly makes
the output any more substantive.
It includes phrases like "not every castle is old", "not every day is old", or
"not every tower is large", or "not every look is free".
Lutz then printed selections in Benze's literary magazine Augenblick in 1959.
The stochastic texts, as I said, were one of the first examples of natural
language processing in Germany, and they proved that computers could operate
not only on mathematical symbols but also on language.
They were also artificial poetry in Benze's sense.
No matter how many variations the program churns out, there is no ego
expressing itself and no consciousness standing behind it all, vouching for the
meaning of the words, which are merely concatenated according to weighted
random operations.
That the computer itself could actually be the author of this text seemed
absurd to both Lutz and Benze, but both knew how it had been produced.
Whether its artificial origin can be recognized is less clear.
The readers of the magazine Augenblick were not compelled to ask this question,
and an accompanying essay enlightened them to the details of its creation.
So they knew it was computer-generated.
But when the following year Lutz generated a second poem according to the same
pattern, it was titled "And No Angel Is Beautiful" or "Kein Engel ist schön",
and instead of Kafka had used Christmas vocabulary, and he published it in the
December issue of the youth magazine Yard 9, there was no explanation to be
found.
The poem was placed on page 3 along the miscellanea, just like any other poem.
Only the author's name, Elektronos, might have allowed one to guess who or what
was behind the text.
The next issue solved what most readers had not even identified as a riddle,
namely that a computer had written the poem.
Clearly Lutz was having fun here, as is evident from the ironic captions under
the photo of the author, the Z22, and the second poem in the poet's
handwriting, as Lutz wrote, that is a teletype printout.
On the same page he published a series of letters to the editor.
The writers, without knowing how it had come about, were quite divided in their
assessment of the poem.
"Perhaps you should reconsider whether you want to open the columns of your
paper to such modern poems", poets complained one, while another was on the
contrary impressed by the avant-garde stance.
"Finally, something modern."
A third reader was at least open-minded.
"To be honest, I don't understand your Christmas poem, but somehow I like it
anyway.
One has the impression that there is something behind it."
Evident in these reactions is what I would call the reader's standard
expectation of unknown texts.
The electronist poem was indeed artificial poetry in Bent's definition, a
computer-generated text without meaning, mediated by an authorial consciousness.
But because its readers were unaware of the conditions of its production, they
took it for a natural text and assumed it was written by a human with the aim
of communicating meaning.
The standard expectation of unknown texts, then, can be captured as a
relationship between two elements, which sometimes is extended by a third.
First, that the text has an originator, namely that a human, or sometimes more
than one human, wrote it.
And second, that the text has intentional and semantic content, that a
communicative will to meaning, sometimes understood as reference to a world, is
expressed in it.
In some cases, there is also, third, the text's connection to an author
function, that this constellation can and should be subsumed under the name of
an author which organizes the attribution and circulation of some text, but not
all.
But as central as authorship is for literary studies, I believe the first two
elements are more important in this context, and I will therefore foreground
them.
I take pains at separating these three elements because they are not identical.
The first element, the human, for instance, does not by necessity imply the
second, the intent to communicate, say, in the case of Dionysian writing
frenzies.
Nor does the second by necessity imply the third, the author, since an author
is more than just one intending a text.
Further, both intent and authorship can be uncoupled from the human originator,
because historically we know of texts that were read as not having been brought
about, or authored solely by humans, such as holy books, that express a divine
will, even though the instance recording is usually human.
I think that Moses's tablets are one of the few exceptions.
And while the intentional element may be internally divided in the debate about
the relationship between meaning reference and communicative intent, a big
topic in AI text at the moment, most of these debates ignore that the assumed
originator is more than just a nondescript subject, but often, possibly,
mostly, specifically understood as a human.
One should at least ask whether the assumption of intent alone is enough for
readers to designate something like a computer as an author, since it is often
precisely the simulation of humanness that acts as the assurance of intent in
the first place.
I will come back to this point.
Important for now is that a reader's standard expectation of unknown text is at
minimum this, that it was written by a human who wants to say something.
That there is a standard expectation at all, even if it may be limited to
modernity or Western modernity, has only become apparent since there has been
an alternative to it, and Benz's conceptual distinction and Lutz's practical
demonstration have been illustrative here.
As the incensed letters to the editor show, to recognize a text as violating
the standard, that is, as artificial, always requires additional information,
if it is not provided a human origin is assumed.
In this regard, Lutz had indeed given his readers the runaround, as one letter
to the editor complained, not because a modern poet had written bad but natural
poetry, but because a computer had generated meaningless because artificial
text.
Passing of an artificial text as a natural one was not just the debut of a
rather hackneyed joke made by the computer scientist in a provincial youth
magazine in the 1960s.
On the contrary, this runaround is the principle of artificial intelligence,
and at the same time that which connects language technologies with an imputed
humanness as an element of the standard expectation.
Ten years earlier, in an article that became the founding document of
artificial intelligence, the computer science pioneer Alan Turing had pondered
whether computers could ever be intelligent.
Turing famously rejected this question as wrongly posed.
Intelligence as an intrinsic quality could not be reliably measured.
In good behaviorist fashion, he therefore replaced the question with another.
If we assume that intelligence is a property of humans, then all we need to
find out is when humans would consider the computer to be itself human and thus
intelligent.
Note that Turing at no point here speaks of intent, a category he finds
meaningless.
The experiment setup is well known.
Three participants, a human judge, a human respondent and a computer,
communicate solely through text via a teletype printer, with the judge's task
being to determine which respondent is the machine based on their answers.
If the judge cannot reliably distinguish between the human and the machine, the
machine is considered to have demonstrated human-like intelligence in that
context.
Of course, the original setup has a gender component that I'll pass over here.
The point is not that the answers in this conversation have to be correct, but
that they sound human.
Lying and bluffing are explicitly allowed.
If one wants to examine the reader's expectations of artificial text, Turing's
test is still a helpful starting point.
First, because his setup assumes a teletype conversation and thus equates
intelligence with written communication.
And second, because the goal of this communication is to misrepresent signs
that are meaningless to the machine as meaningful to humans.
To put it bluntly, the essence of AI is to pass off artificial texts as natural
ones.
It is only worthwhile to make this attempt at all, however, because the
standard expectation of unknown text is that of human origin.
Artificial intelligence, then, as a project, if not in each of its actual
instances, is based on the principle of deception from the start.
For this reason, media scholar Simone Natale writes, "Deception is as central
to AI's functioning as the circuits, software and data that make it run."
The goal of AI research, she says, is "the creation not of intelligent beings,
but of technologies that humans perceive as intelligent."
With an eye to Turing, one might add, as close as possible to being human.
I would like to call this position strong deception.
Yet because this conception of AI is essentially deceitful, we can ask whether
the expectations of a text could ever change under these conditions.
I think not.
It insists that artificial and natural texts remain neatly separated so that
one can be mistaken for the other.
If it is suddenly revealed that a natural text is in fact an artificial one,
its readers will feel cheated, and not without reason.
Deception here turns into disappointment.
Deception turns into disappointment.
We don't know how Theo Lutz's readers reacted to the revelation that the
computer had written the poem, but one can guess if one considers recent cases
in which the artist subsequently turned out to be a machine.
This prominently happened in June 2022 at a rather peripheral art prize.
When a participant admitted that he had not painted his entry himself, but that
it had been generated by the text-to-image AI Dalí, a torrent of indignation
followed, and he was accused of fraud.
Even though this was an art prize for digital art, this apparently referred
only to the tools.
The artist himself was still supposed to be human.
Again, it is the originator, not the presumed intent, that was at issue here.
A somewhat similar case occurred in Japan, where the author Riku Dan, winner of
the 2024 Akutawaga Prize for Literature, disclosed in an interview that she had
generated portions of a prize-winning novel using Charged GPT.
Again, public scorn followed, and she was accused of fraud.
There are other such examples, and although these disappointed expectations are
usually exaggerated in the press, they reveal what was actually expected,
namely natural, not artificial, text.
However, neither the neat separation of Benz's ideal types, nor this
expectation itself can remain intact as soon as the exception becomes the rule,
that is, as soon as we are surrounded by text whose origin is unclear.
At first glance, such examples seem to suggest that the reader's expectations
of unknown text have actually not changed since Lutz's time.
We assume human origins and communicative intent, which is why deception can be
a useful strategy in AI design in the first place.
But in fact, I believe that expectations are nevertheless already in the
process of shifting, and this has become even clearer in the last two years.
Because the number of computer-generated text is constantly increasing, and
because we ourselves are writing ever more with, alongside and through language
technologies, we are on the way to a new expectation, or rather a new doubt.
The more artificial text there is, the more the standard dissolves and the
question of the origins must arise, even when we normally would not think about
it at all.
That there nevertheless has been a shift can be explained by the fact that the
examples of text I have considered so far are special ones, they are literary
texts, texts that are marked as exceptional in our cultural tradition.
They appear to be intended and worked through to the smallest detail, and they,
more than other texts, are read as having an author who is human, who wants to
say something.
And yes, this is still true even after 60 years of talk of the death of the
author.
This also has been a lesson of the last few years.
It is worth taking a look at the other side of the spectrum first, however, at
those rather unmarked automated texts that remain in the background, that are
merely functional and that do not assert themselves as products of a human
intent or a strong notion of authorship.
For them, the Turing test is simply a false description of reality, and the
standard expectation is already in the process of fraying.
For there are forms of human-machine interaction other than strong deception
and other text types than the artificial-natural partition would suggest.
Especially when engaging with interfaces, we are likely to find ourselves in an
intermediate stage between natural and artificial.
Here already, we can experience a looming shift in the standard expectation.
It is quite possible to know that something has been produced by a
non-intelligent machine, at the same time treated as if it were conscious
communication.
In fact, this is quite normal.
Natale has proposed the term "banal deception" for this phenomenon.
In contrast to what I have called "strong deception", here users are aware that
they are being deceived.
We understand that Siri is not human and does not have an inner knife, but
smooth communication with her works only if we treat her, at least to some
extent, as if she had one.
Knowing this is not a contradiction that suddenly and unexpectedly destroys an
illusion, as in the examples of competitions in which an AI participates
surreptitiously.
Instead, a banal deception becomes a condition of functionality.
If I do not play along, Siri will just not do what I want.
The situation is very similar with written text.
It starts with a dialog box on the computer screen like this.
After all, the question "Do you want to save your changes?"
enables an interaction that is basically similar to the one with a human being.
The answer "yes" has a different effect than the answer "no", and both lie on a
continuum of meaning that connects natural language with data processing,
without one suspecting any meaningfully strong notions of intent behind it.
Something like this would already lower the expectations of unmarked text.
While we still act as if we expect human meaning and a conscious interest in
communication, we bracket the conviction that there really must be such things
involved.
This bracketing means discarding the third element, namely authorship, while
entertaining the second element, intent, in an assertive modality, and possibly
the first, human origin, in a fictional one.
Yet this bracket does not always proceed smoothly.
Banal deception is an "as if" that demands of us the ability to hold a
conviction and its opposite simultaneously.
This self-contradictory position quickly gives rise to a doubt.
The more convincing artificial texts become, and the more the aesthetic
impression they make on us suggests something like human-like intent, the more
difficult does it become to feel comfortable in the limbo into which banal
deception lures us.
If artificial texts become indistinguishable from natural ones, and if moreover
we know that computers are capable of writing them, a new standard expectation
of unknown texts lies before us.
It is the doubt about their origin.
Rather than taking a human source for granted, or simply deferring the
question, the first thing we would want to know about a text would be, was it
made by a human or a machine?
With the increasing integration of large language models into existing
software, Google Docs incorporates Gemini-based assistants now, Microsoft Word
uses Chat GPT, it becomes increasingly difficult for readers to clearly
classify such texts as either human-made or machine-generated.
The stakes seem relatively low when it comes to LLM-written marketing pros, but
what about the lawyer's letter that might be automatically generated even
though it's about my own personal case?
What about my students' essays that I have to grade?
What about political articles or fake news stories?
What about the private, personal, intimate email, the love letter?
Are those AI products too, in whole or in part?
At least one reason for the discomfort these ideas evoke is that people have a
stake in what they write and to varying degrees, they vouch for their words.
While scholars of literature have learned to read without an eye to what an
author wants to say, and merely regard the semiotic interplay of signifiers,
this is still the mode in which most everyone else reads any written document.
Even if a text ultimately turns out to be inaccurate or misleading, the
standard expectation that a recipient brings to reading it involves the
assumption that the author is making what Jürgen Habermas has called "validity
claims", among which the validity claim to truthfulness or sincerity is maybe
the most relevant in this context.
Essentially, it means that we have a basic level of trust that speakers or
writers mean what they say rather than try to deceive.
This is the reason that reading critically has to be learned at all.
Whether or not readers ultimately judge a text's assertions to be true, they
tend to assume the existence of a writer who does.
That's the difference between truthfulness and truth.
If truthfulness is thrown into calamity once large language models can generate
texts that appear to have been produced and sanctioned by an author, the same
can be said of truth, that is, correctness, another of Habermas's validity
claims made in speech acts.
We know that LLMs are still lacking when it comes to the handling of knowledge,
understood as reporting pre-established facts or correct data, if knowledge is
simply the probability distribution of tokens over training data.
They may merely learn the form of a specific genre without any scientific
insight, responsibility or accountability.
Among many examples of LLM-made scientific papers, I still find the
Chan-GPT-generated legal brief most striking, which referred to non-existent
laws and cases filed by a Manhattan lawyer in 2023, which would have fooled
anyone not familiar with the law in question.
And even them, as this lawyer aptly demonstrated.
The text-pocalypse, then, as Matthew Kirschenbaum is calling it, the
proliferation of generated text, is also a crisis of truth and truthfulness.
To maintain the standard expectation, with its separation of natural and
artificial text and its assumption of human origin to tend to end for
especially marked text authorship, is a challenge under such circumstances, to
say the least.
Given the crisis of the standard expectation, it is not unreasonable to suggest
that it is already shifting from the conviction that a human is behind a text
to the doubt of whether it might not be a machine after all.
But this would also make the distinction between natural and artificial text
increasingly obsolete.
We would then possibly enter a phase of post-artificial text.
By this I mean two related but distinct phenomena.
On the one hand, post-artificial refers to the increasing blending and blurring
of natural and artificial text.
Of course, even before large language models, no text was truly natural.
Not only can the mathematical distribution of characters on the page, as Benzer
had in mind, also be achieved by hand, but it is the truism of media studies
that every writing tool, from the quill to the pen to the word processor,
leaves its mark on what it produces.
On the other hand, no text is ever completely artificial.
That would require real autonomy and actually strong AI that could ultimately
decide for itself to declare a text published.
Today, however, with AI language technologies penetrating every nook and cranny
of our writing processes, a new quality of blending has been achieved.
To an unprecedented and almost indissoluble degree, we are integrating
artificial text with natural text.
In the wake of large language models, it is not implausible that the two types
of text might enter into a mutually dependent circular process that
irreversibly entangles them.
Since a language model learns by being trained on large amounts of text, so
far, more text always means better performance.
Thinking this through to the end, a future monumental language will, in the
extreme case, have been trained with all available language.
According to one study, this may happen already in the next few years.
Let's call it the last model.
Every artificial text generated with this last model would then also have been
created on the basis of every natural text.
At this point, all language history must grind to a halt, as the natural
linguistic resources for model training would have been exhausted.
This is to Ouroboros' problem.
LLMs trained on synthetic text, the only text for it to swallow, suffer from
sudden degeneration, a phenomenon called model collapse or autophagy.
More importantly, the language standard thus attained would in turn have an
effect on human speakers again.
It would have the status of a binding norm, integrated into all the mechanisms
of writing that build on this technology, and which would be statistically
almost impossible to escape.
Any linguistic innovation, any new word, and every grammatical quirk that might
occur in human language would have such a small sham in the training data of
future models, that it would be averaged out and leave virtually no trace.
This is of course a deliberately exaggerated scenario.
As a thought experiment, however, it shows what post-artificial text might be
in the extreme case.
But even before that happens, halfway to the eschaton of absolute blending or
erasure of natural and artificial language, a new standard expectation of
unknown text might already emerge.
This is the other meaning of post-artificial and the one I am primarily
concerned with here.
After first the tacit assumption of the human origin of a text, and second the
doubt about its origin, it would be the third expectation of unknown text.
For doubt about the origin of a text, like any doubt, cannot be permanent.
Humans have an interest in establishing normalcy and reducing complexity and
uncertainty to tolerable levels.
Already mechanisms are put into place to keep the doubt in check, either by
assurance or by decree, by digital certificates, watermarks or other security
techniques designed to increase confidence that the text at hand is not just
plausible nonsense or by laws.
However, to quote Hegel, one bare assurance is worth just as much as another,
and the law does not abolish the crime.
The fact is that technical checks can always be circumvented, and there is
currently no surefire way to detect AI-generated text.
In fact, there are good reasons to believe that none is possible in principle.
What this means, however, is that once it is possible to question whether a
text might have been generated by a machine rather than written by a human, no
certificate or law can extinguish that doubt.
If it can neither be resolved nor, as I believe, borne permanently, the only
solution is to undo its premises.
Should political regulation and technical containment fail, then, it is not
unlikely that the standard expectation itself will become post-artificial.
This is the second use of the term.
Instead of suspecting a human behind a text, or being haunted by scepticism as
to whether it was not a machine after all, we simply lose interest in the
question.
We might then focus only on what the text says, not who wrote it.
Post-artificial text would be agnostic about the origin.
If the standard expectation of unknown text is shifting, if it is increasingly
riddled with doubt, perhaps even capitulating to such an agnostic position, why
the ostentatious excitement over generated text in literary competitions?
This is, I think, because literature is slower than other forms of text.
And this is because, beneath the notwithstanding of all text types, it makes
the most emphatic claim to a willful human origin, all the while connecting it
more forcefully than any other type of text to the historically grown notion of
authorship.
I have already said that there are texts today whose origins do not pose a
question.
They are unmarked.
A street sign has no author.
In our daily life, a news site's weather forecast is also practically
authorless.
Until now, however, we have always assumed that the human being is behind it.
But under post-artificial reading conditions, nothing much changes if we simply
make no assumptions at all.
My belief is that more and more texts will soon be received this way.
Put differently, the zone of unmarked texts is expanding.
Not only street signs, but also maybe blog entries, not only weather forecasts,
but also information brochures, discussions of Netflix series, and even entire
newspaper articles will tend to be unmarked.
And it is not unlikely that they too become writerless and in many cases
authorless.
Literary texts, on the other hand, are still maximally marked today.
We read them very differently than other types of text.
Among other things, we continue to assume that they have not only a human
writer, but also an author.
The consequence of this markedness is that art and literature themselves have
recently become the target of the tech industry, namely as a benchmark to be
used after other formerly purely human domains such as games like chess or Go
have been cracked.
Now art and literature pose the latest yardstick.
Probably nothing would prove the performance of AI models better than a
convincingly generated novel.
Ultimately, however, this hope is still based on the paradigm of strong
deception.
Indeed, there is currently a whole spate of literary and artistic Turing tests
to be observed that all ask, can subjects distinguish the real image from the
artificial one, the real poem from the AI-generated one?
These tests mostly come from computer science, which as an engineering
discipline likes to have metrics to measure the success of its tasks.
The problem is that they still compare the rigid difference between natural
expectation and artificial reality.
This seems to me of little use when it's this difference itself that is at
issue.
More interesting then is the question of the circumstances under which this
difference becomes irrelevant.
In other words, what would have to happen for literature to become
post-artificial?
I will close by briefly trying to sketch an answer to this question and by
returning to the standardisation tendency that arises from the run towards the
last model.
In them, a normalisation takes place.
The outputs are most convincing precisely when they spew out what is expected,
what is average, what is statistically probable.
The more ordinary a writing task, the more easily it can be accomplished by AI
language technologies.
And just as marketing AIs now assist in the creation of marketing pros that
meets our expectations for marketing pros, there also are literature AIs that
assist in writing predictable literature, as one might say.
Predictability, as that what can be expected, may be described statistically as
a probability distribution over a set of elements or, on a higher level, a set
of patterns.
The more recurrent they are, the more likely and expectable the outcome.
One popular prediction, and one that seems plausible to me, is that genre
literature, which is virtually defined by the recurrence of certain elements,
is particularly suitable for AI generation.
A banal famous case is that of Jennifer Lepp, who writes fantasy novels under
the pseudonym Leanne Leeds, like at an assembly line, one every 49 days.
In this process, she is aided by the program Pseudowrite, a GPT-based literary
writing assistant that continues dialogues, adds descriptions, rewrites entire
paragraphs, and even provides feedback on human writing.
The quality of this output is quite high, insofar as its contents are just
expectable.
Since most idiosyncrasies are averaged out in the mass of training data, they
tend toward a conventional treatment of language within the bounds of a certain
genre.
They become ouroborous literature themselves, maybe.
At the moment, machine learning is not enough to generate entire novels, but I
do not see why just this kind of literature could not be produced in an almost
fully automated way very soon.
Maybe aided by human acting as an art director or series editor, reducing the
49 days to 49 minutes or even less.
If the prediction is allowed, I think it would be this kind of literature that
is most likely to become post-artificial.
Of course, author names would not disappear, but they would function more as
brands, representing a particular tested style.
Just like some books series today are written entirely by committee, but which
we still assume to be a collaboration between humans like Tom Clancy or
something like this.
The unmarked zone would extend to certain areas of literature, not all, and
certainly not all narrative ones, but far more than it encompasses today.
Conversely, we might ask what kind of literature is most likely to escape this
expansion.
Here I see two answers that at first glance seem contradictory.
If the unmarked post-artificial literature is one that absolutely mixes natural
and artificial text, then writing that clings to this marking would be one that
emphasizes their separation.
On the one hand, then, one could imagine the emphasis on human origins as a
special feature.
Again, ex negativo, we can already observe phenomena that point to such a
development.
On the web, for instance, artists have been up in arms against image-generating
AI, such as Dali or stable diffusion.
They recognize stylistic features of their own work in the generated output,
which may therefore have been part of the training set.
This raises very legitimate questions about copyright and fair compensation, a
discussion that is unresolved and ongoing.
At the same time, however, there is also resistance to AR generated per se,
which some fear threatens to make human artists obsolete.
On social media, the hashtag #SupportHumanArtists has emerged as a declaration
of war against generative image AI.
One can imagine something similar for literature.
Perhaps even a future in which the label "guaranteed human-made" is considered
to be a distinction.
Just as one buys handmade goods on Etsy, one can imagine a kind of boutique
writing that boasts of its human origin as a proof of quality and as a selling
point.
Such re-humanization could boost certain genres of literature, such as the now
popular auto-fiction.
Playing with the identity of author and narrator, auto-fiction insists on the
human origins of a text.
The same goes for experience-based non-fiction, such as the memoir or the
personal essay.
They too make the human behind the text, the condition of its reception, and
are thus especially suited to confronting the post-artificial situation.
But if one does not want to rely solely on an external assurance of human
origins, which in any case still leaves room for doubt, which is in principle
impossible to dispel, an unpredictable, unconventional use of language can
indicate writing beyond the model's abilities.
Every formal experiment, every linguistic subversion, would oppose the
homogenizing probability of great language models, their leveling Ouroboros
standard.
Linguistic unpredictability would then be evidence of human origin.
In the most extreme case, the sign system in which language AIs operate would
be exploded, as in the case of visual and isamic literature, like that of
Kristen Miller.
She no longer uses any letters at all, but only the impression of lines and
blocks of text.
The pure poetry Max Benze dreamed of would paradoxically not come from the
machine, which now, in post-artificial blending, plausibly simulates meaning,
but from people who no longer attempt to make sense at all.
Thus understood, and this is the second route of highlighting a text as non-AI
made, the descendants of Lutz and Benze at least have a chance of escaping the
post-artificial situation by continuing to mark the artificiality of their
products.
This is digital literature, literature that is self-reflexively produced with
the help of computers.
It can escape the post-artificial, at least to some degree, by consciously
emphasizing the entanglement between the natural and the artificial, rather
than glossing it over for a natural-seeming appearance.
Much more than conventional writing, digital literature always keeps a critical
eye on its origins.
One example was Mathias Kuhn's German-language book Selbstgespräche mit einer
KI, or Monologues with an AI, in which, in addition to his literary
experiments, Kuhn also provides the source code for training the language model
and even its training data, which is a very small sliver of text.
The human and machine components that together produce text can, not
completely, but at least somewhat, be separated here.
Conversely, a deliberately staged human-machine collaboration can also have
this analytic effect.
In David J.
Johnson's rewrites from as early as 2017, the author trained the language model
every night for a year and then edited the output by hand the next morning, in
a process he calls "calming".
The point at which the machine hands over its text to the human jarv is
precisely marked, and by collecting the edited results of each month in a book,
so that rewrites now comprise 12 heavy volumes, he also frames this
collaborative, but not absolutely fused process as a performance, which is also
not conventionally literary.
Of course, no proof of human intervention is ultimately provided here either.
But perhaps the obstacles that can still be put in the way of the
all-too-smooth reception process is the maximum of resistance to the
post-artificial that will still be possible, before the difference between
natural and artificial has disappeared altogether.
It should have become clear that I have entered highly speculative territory
here.
I am not suggesting that narrative or, broadly speaking, conventional
literature is now doomed, or that only experimental and explicitly digital
literature is worth pursuing.
Nor do I mean to imply that post-artificial texts are necessarily bad.
One can certainly enjoy reading them, discuss their merits, and unravel their
interpretive dimensions.
Here, I have been primarily interested in analysing tendencies.
And for this purpose, it is worthwhile to consider possible extremes.
As far as literature is concerned, there is of course a third possibility, that
AI may, through some technological innovation, be steered to produce less
probable and more interesting output, without losing all the advantages that
the power of normalisation provides in coherence and meaning production.
It is a matter of optimism, or maybe lack thereof, to consider this a likely or
unlikely future.
Seeing that sufficiently large LLMs are still tied to capital interests, I
remain apprehensive.
In this talk, I wanted to think about how language is changing in the technical
age we inhabit today, and which will continue to unfold, both without fearing
the technology, but also refusing to succumb to its ideologies.
In that context, one thing seems certain to me.
With the increased penetration of language technologies, with the triumph of AI
models, our expectations as readers will change.
Thank you.
Thank you so much, Hannes.
That's a very stimulating, disturbing perspective.
I am currently beginning a book on what I would call the second death of the
author, after the first death of the author, claimed by the structuralism area,
when they decided that because there were so many texts under a new one, the
history of language and the history of literature was speaking instead of the
actual author.
I would be very interested to read the paper you certainly have about it and to
quote it in my research.
I have many questions.
As you know, I am a scholar in literature.
I have a book on the history of the idea of literature, so it's very disturbing.
Because when you have literature, as you said, you have a lot of values coming
together, like authenticity, like expressivity, and maybe not just the value of
the truth of what is said, but all that comes with the idea of having someone
communicating with you.
There is a very famous quote by Blaise Pascal saying, "You are looking for an
author, but in fact you are looking for someone, for a man."
Maybe behind the abstraction of authorship, what we are looking in a book is a
kind of authenticity of experience in comparison to what we got in electronic
literature.
For instance, in France, about 50% of the books published are written in the
first person, because they claim authenticity, they claim to give an account of
oneself.
To quote Judith Butler, something that is very situated, that is anchored in
something that is ground in some experience of life.
So I think there will be a lot of resistance to completely artificial
authorship, but I don't want to monopolize, and I give the floor to everyone,
because what you propose is really stimulating.
It is the same case in the case of photography, because we no longer presume
that photography is a kind of reality, like Roland Barthes wanted to say, when
you get photography, you get a piece of reality.
It's not true anymore, so what happens in photography is a kind of
post-artificial reading position.
Maybe I can directly answer this, because I think this is an interesting
question of photography, CGI, post-truth, the treachery of images and so on.
But I believe this is exactly not the same case with post-artificial writing,
because in the case of computer-generated images and photography, we are
talking about the depiction of reality or the representation of reality is a
question.
A text is not a representation of reality in the same way that a photograph,
which comes through a lens, is an index of an outside world and then is
received by some kind of surface.
So I think the difference here may be one between noises and noema, in the
Rossellian sense, between the object of consciousness and the way of giving
this to consciousness.
In other words, in a photograph, the noema, the object, is at issue.
It could be false or true.
And in post-artificiality, I think it's rather the givenness itself, namely
whether this is an expression of the content of a human mind or not.
And I think that's a difference in structure to me, because when we write
something down, we don't represent reality.
We can represent all kinds of things.
It can be fiction.
But the way it is given through the mediation of a human author has this mark
or the origin of the human inscribed in that.
So I would just take pains to see what the differences between these two media
are, because I think they are somehow different.
I understand.
I understand.
What does the public think about all of it?
Let's listen to possible questions.
I had a remark, maybe.
Can you hear me?
Yes, absolutely.
Yes, thank you for your very stimulating presentation.
I was thinking while listening about maybe other examples of not text, but
stories where authorship doesn't matter.
And I thought about two or three examples.
First, maybe stories for children, where they don't seem to put any importance
on whether who is the author and what was the intention of the author.
It's more a matter of whether the stories works or not, whether they are
grasped and they feel immersed in the story or not.
So that was the first example.
I was thinking also about stories in oral societies where they are passed from
generations to generations with small variations, but where there is no clear
author and in the same way, there is no real importance on any intention of
someone on any subjectivity, but rather if it works or something else is at
stake, I would say, like related to traditions and relation to the environment
and so on.
Where even you could have non-human authors.
I was thinking about the Australian tradition where the author of the dreams is
more the land and the characters are part of the land.
And it's not a human subjectivity.
So all these examples made me think maybe we are overestimating the importance
of authorship because we are scholars and educated people who put a strong
emphasis on the context of the text.
And what's the subjectivity that has been expressed or adults in search of
shared subjectivities.
But all these examples maybe put another emphasis.
So I just wanted to suggest these examples and see how you react.
I think these are great.
This is a very good idea.
And stories for children seems to me the one that makes the most sense.
Because oral societies can have authors in kind of a mythical way.
The whole idea of Homer being a person might be a retro projection from a
society that needs authors, but he was clearly already a name in circulation.
Although myth would be another category.
But myth is something that is in circulation not simply because there is no
author but because there is no beginning to it.
No temporal beginning in time to myth.
I like the idea of children's stories particularly because that distinction
that authorship introduces between reality and fiction, believability or
something like this, doesn't seem to play a role.
And maybe one could then make a polemical argument that there is a certain type
of infantilization coming with the idea of not having an author.
I wouldn't necessarily believe that because there is a distinction between the
possibility of verification, if authorship is some kind of legitimation
verification of the truth of what is said.
Post-artificiality does not think that we can completely do away with it, but
it has no other.
It is kind of a giving up.
It is a situation in which this is not possible, so we have to search for other
strategies of dealing with that.
Which would be slightly different than the situation of a children's novel
where this is never a question.
The verification of the truth of the matter is not an issue at all.
So I would think that there might still be a difference between those.
But I do like the reminder that these exist and that there are modes of reading
that obviously don't need authors or don't need truth to that degree.
Maybe to elaborate the interesting aspect in children's books is the efficiency
of the story.
And I was wondering, do we have examples of stories that are very efficient in
a way that they are compelling and people want to read what's next and would
have been artificially produced?
I don't see any reason why it would not be possible to have compelling stories
with no authors.
So where the focus would be more on the… The best word I can find is efficiency.
I'm not sure it's the best.
The efficiency of the story, like something intrinsic in the structure that
makes it compelling.
I know, I mean, as I said, I don't think there is a fully, really fully
computer generated novel that exists that would fulfill this criterion.
But a couple of years, I think by now ago, there was a slight scandal about the
circulation of computer generated children's cartoons on YouTube.
There was like a whole, like a wave of these, which are to a degree almost
completely incoherent.
They are cut together from different bits of… are generated in a way that, you
know, if an adult would watch them, it just wouldn't make sense to them.
But apparently, because a lot of parents put their, you know, give their child
an iPad and just play the lists of all, you know, Peppa Pig or whatever it's
called, going through.
This was apparently quite successful because, you know, it was watched a lot.
I mean, we can think what that does to children if this kind of incoherence is
the basis of learning how stories work.
But clearly, it seems that as a genre, there's something else to it than simply
efficiency or compellingness.
It can also simply be the draw of it going on and on and on.
It might be the medium of the cartoon or the film rather than text that has to
be read by a parent or something like this.
But all I wanted to say is, especially for children's media, there seems to be
AI generated stuff, which is worrying, I believe.
I also have a question, if I may.
Yes, of course.
I totally get your points on photography and how it is different from LLM's.
And I agree with you.
However, forgive me, my question is going to be related to photography.
And it's regarding when you talked about deception and disappointment.
I'm very interested in the affective relationship we have with tools.
And I was thinking, because when I was a bachelor student, I remember this
story one teacher told us.
And I actually never found a decent book about that.
So I would be happy to find it one day.
But I remember she said that actually, photography wasn't taken seriously by
military in the 19th century.
And actually, photography served as a propaganda more to illustrate the
battlefields.
And so it wasn't a representation of reality.
It wasn't something in which the military trusted.
They trusted more the drawings than the photography.
Today, it would seem very crazy to say that.
But they had time to adapt and to actually choose the photography instead of
the drawing.
And maybe we had the same story with drones.
I'm not sure.
I don't know anything about military history.
But it made me think also of the relationship we have nowadays with Google.
And how sometimes people say, "Oh, I googled this."
So it's true because I googled this.
Don't you think one day the effective relationship, what I mean by that is the
trust we have in the tool, might change from deception to full trust?
This is a very optimistic vision of it.
Yeah.
That's a good question.
I think there would be another possibility to think through that thought
experiment.
I'm not sure, but it seems to me that there are reasons why we give trust and
don't.
I'm not a scholar of photography, but I believe the original way it was
described by Niepce, Fox Talbot and people like this is that it is an artless
art.
It is one in which human intervention no longer plays a role.
One way to say this was that it's the pencil of nature.
The pencil of nature is drawing the photograph.
I think that the direct indexical relationship between what is out there and
what is on the paper was the starting point.
And then it became more convoluted because then people noticed that while there
might be truth to this, obviously photographs can lie.
Even before they are manipulated, you can just simply propaganda the use of
photographs and so on.
I'm not sure if we have the same starting point for this parallel.
Because the whole point of large language models is the statistical production
of language, which is from the start not concerned with truth or factual
correctness.
So to get to a position of trust to this again when you start from non-trust,
which is the reverse of photography, where you start with trust and then you
have to learn to not trust it.
I wonder what would have to happen for this.
And I think what would have to happen is all kinds of mechanisms of
verification, giving credence to the statements, double checking things.
One thing people are interested in now is this RAG technology, the idea that
you can plug a large language model next to a database.
And the idea is that the language model would formulate the question, translate
the question into a search query in SQL or some other kind of database
retrieval language.
And then it could retrieve something that is by form as a database, somewhat
verifiable and repeatable.
This might be one way to go, but I still think we have big problems with this.
This is why you always see under Cloud, I used the new version of Acrobat
Reader the other day, which now is an AI function.
You can ask it questions about the PDF you're reading.
It always says, please double check.
And I don't know if we can get away from this, because all the technological
precedent I can think of in which we start with non-trust and go to trust.
One example would be Wikipedia, where when I was in high school, everybody
stressed that you cannot trust it.
And it's not a source.
And I think we're not there anymore.
And this was achieved through basically the implementation of very complex
policies and structures and processes to ensure verifiability, veracity and so
on.
So obviously everyone is working on this, right?
All AI companies have a vested interest in making the technology more
trustworthy.
But I'm not convinced that this is possible in the long run.
Obviously, this whole argument that once a doubt is in the world, you cannot
extinguish it anymore is one made from extremes.
We will find probably some kind of modus operandi that allows us to work with
these things.
But really, the trust in the strong case, I don't know, seems unlikely to be
reinstated in some way.
Thank you very much.
Can I ask a question?
Yes.
So I'm Yanis, I am a computer vision researcher and I have done a bit of early
AI art before ChessGPT when there was GPT-2, the first LLM that kind of worked.
And there was a version that a lot of artists used that was called the
SteelGPT-2 that was very easy to install in a normal computer.
And I remember we made the book and I gave it to a French person and the way
that this French person reacted was, this is not French.
So he said basically that we had French people on the book.
I didn't know well French at the time, but the book was edited by French people.
So it was French in a way, but he reacted basically in a subjective thing.
So he reacted on the originality on the book.
So the book had a certain subjectivity and early models had subjectivity.
They gave this thing, this feeling that it was a person.
But what they lacked is that they lacked a certain community.
They gave you something that you didn't really want or they didn't want you to
make a community with them.
And with models now, we have another thing which is a bit like this
subjectification.
So basically right now we have something that reminds us of the writing that
comes from computer science, which is a very different type of writing.
That is the writing that comes from the humanities where you have to totally
destroy your subject.
Whoever writes that should be the same.
Maybe some types of writing should be better in the sense of how you transmit
information, but your subject should kind of vanish.
And in both cases, we have a form of a non-psychological form of the machine in
the sense of also Baudrillard that a machine has no other.
So the machine either is very narcissistic, very into itself.
In the other case, it's just a tool, something completely that operates on a
certain task.
And the question there arises on it's very clear that these models right now
are not made for writers.
They are not made for people that write.
They are made to write.
So it's and they are not created with writers.
But the very good question, which is not the case, for example, with AI art
that has to do with making images is that writing is very accessible.
Everybody can write.
Everybody can learn to write.
We can argue on how much this is true, but it's much more accessible, let's
say, than doing a painting or doing other things.
And the question is, do writers really need that?
And the way that writers are integrated in this pipeline, because it feels like
something that is disconnected from writing and writers can integrate in their
pipeline or not, but is a bit disconnected.
And what is your take on that?
What is the relation of this tool with writing?
Because it feels like it's it feels like people that use this tool mostly don't
write and don't want to write.
And they use this tool to write or they don't want to spend time writing
because it's a topic, of course.
But so what is it?
I think this is true because this is also how it is being marketed.
Right.
You you can it's usually JGPT is even though there's a lot of buzz about the
notion of creativity, that is not the that is not the use case.
The use case is the production of the increasing effectivity.
Right.
And if you if you write in your job, you have to write legal briefs.
That's why that lawyer is interesting.
It makes it easier, but it also makes more perilous because you might it might
hallucinate.
So, yes, it's not made for for for writers.
And it is it is kind of a subsystem that can be implemented in all kinds of
professional writing tasks in which writing is not a.
Is it communicates a certain.
Basic functions like, you know, if you write professional letters or something
like this, this is not it doesn't draw it.
It doesn't have a poetic.
It doesn't have a poetic dimension.
It doesn't draw attention to itself.
However, someone like, you know, what I what I what I showed you, Leanne Leeds
and the software pseudowrite are meant to augment writing.
And this is interesting because it's not a complete generation, but it is a
process of increasing the production of text also from a from an efficiency
standpoint, but for writers in mind.
So what you can do in pseudowrite is you can ask it to suggest the start, the
start of the paragraph.
You can ask it to rewrite a given paragraph.
You can ask it to, you know, the little buttons for different sensory modes.
So you can say you can select a paragraph and ask, add more visual description,
add more auditory description, add more metaphors, add more, you know, and so
on and keep writing.
So there is there is kind of the idea that even the process of writing
literature is one that can be streamlined to efficiency by giving you these
little shortcuts.
And this does become, I think, relevant as soon as writing is not understood as
an art form or some kind of production of inherent value, but rather as a as a
product that that can be sold.
And I think in that, Lien leads is a very good example because exactly the
production of more texts in shorter times that are sold then to an audience
that that wants these relatively predictable texts that basically are very
similar to each other is simply an economic argument.
And I think the distinction between true writing that has all these kind of
idealistic assumptions and writing as a profession in which you simply have to
produce text in order to live is maybe one that's also from our perspective is
not so tenable.
And maybe the idea of the sociology of writing that takes these kind of
subsistence questions and economic questions more into account would be helpful
here to see that is maybe not immune to this.
But in a way, you cannot say that someone has to be a novelist.
You can say that someone has to write a legal document, but you cannot say that
someone has it's never except if I don't know your father or your ancestors put
you a lot of weight on because of their name or whatever.
But nobody really has to be a novelist.
It's definitely something that should come or comes.
But what you say kind of reminds me of ghostwriting.
So all these people that are paid to not put their name and kind of so it's
kind of automation of this already, let's say, mechanical Turk automated
writing work.
But it doesn't seem like it's a tool that because the thing that really lacks
is this creation of audience.
And unfortunately, writing and a lot of art is in a way, anonymous.
It kind of has to do with it only makes sense for making communities for
audience for the writer themselves.
It's kind of a very psychological need.
It's not like a material need in the sense of solving a problem or solving.
It's a material need in the psychological sense, but not in how to say like,
let's say, yeah, it has to do with humans.
Yeah, maybe.
But as soon as these communities exist, and as soon as a precedent for a
certain text type exists, and especially as soon as there is a market for it,
and there is a market for specific segments, right?
I mean, if you look at the Amazon bestseller charts, they don't have like high
literature, but they have a lot of easily consumable mass produced literature.
And that is something where you can choose to get into this career.
I completely agree, it's probably not a good idea.
And it's to write literature in order to earn money.
But it's still possible.
And I think for that, the automation argument is really one of efficiency
increasing, particularly because and I think this maybe comes back to this
Urboro's argument, we do not start from zero, right?
This transition into a different technical reality is on top of what we already
have.
So we have those communities of readers, we have those precedents of style, and
what a specific type of text is supposed to look like.
And from there on, we suddenly have this kind of increasing acceleration of
automation.
And that maybe is this what produces this odd situation in the first place.
I would agree that maybe this wouldn't be possible if we started at zero, but
we don't, right?
Yes.
Anyway, thank you so much.
Oh, and the other thing I wanted to say, there's this field of critical eye
studies emerging at the moment.
So how can we read these texts?
How can we kind of situate them in their social, economic, political context,
and so on?
And something that I hear over and over again, and I think it's worth keeping
that in mind, Fabien Offert has said that probabilistic systems require a
probabilistic way of reading.
In other words, the single output of a model, be it gRGBT or GBT2 or whatever,
doesn't really tell us much about that model.
Because a single sample is not enough to give you an idea of what this model is
capable of because it is probabilistic.
This is different with something like Lutz and the kind of purely rule-based
systems because there, if you have two outputs, you can analyze them and see
what the structure of this is.
But the breadth and variety of possible outputs just from one or two samples
doesn't cut it in gRGBT.
So I would always say what we as professional readers have to do is to develop
a sense of how one can read these things, given that a single product is not
enough to really form an opinion of their capabilities.
And this is a very practical question.
Thank you so much.
I think if there is no more questions, it's time to thank you so much.
I hope to meet you in person because I would like to have a full-length
discussion with you.
Thank you.
Did you write some papers on which you based your presentation today?
Could you share it?
Yes, this is published.
Actually, it came out in the same issue in Poetics Today as yours.
Okay, it is in Poetics Today.
Because I have yet to read the full issue.
So thank you for the reference and have a good day in California and next time
in Paris.
Bye-bye.
Thank you so much.
Bye.
Bye.
Bye.
Merci à tous.
