Good afternoon, everyone. A very warm and we need a warm welcome today to the National Library of Australia and to this special event. I'm Alison Dellit, the assistant director general for collaboration at the library. Thank you for
attending this event whether you are doing so in person or whether you are joining us from around the country digitally. I'd like to start by acknowledging Australia's First Nations peoples, the first Australians as the traditional owners and custodians of this land. I'll pay my respects to their Elders past and present and through them to all Australian Aboriginal and Torres Strait Islander peoples listening today. Today I am speaking on unceded Ngunnawal and Ngambri lands. This afternoon's programme of events, To Be Continued: Lost Literature and Textual Technologies, is co-hosted by Engaged ANU and the National Library. It's part of
Uncharted Territory, a new ACT Arts and Innovation Festival, celebrating creativity, experimentation, and groundbreaking ideas. I head up the branch at the library that looks after Trove. Trove is more than just something where we reproduce the experience of a library online. It's one of the biggest collections of cultural data available on any single country in the world. So every day we work with massive collection of digital texts, and we're always looking for new ways, opportunities, and risks of interacting with the data that we contain. Today, we launch ANU's new podcast, To Be Continued: Uncovering Lost Literary Fiction from Bushrangers' bushfires to Australian ghost stories, tales of modernity and children's fiction in Australian newspapers that were digitised through Trove.
We will hear from the people behind the podcast and hear live readings from literary discoveries. But before that, we have a very special event. Professor Katherine Bode and some special guests are presenting a panel exploring the textual technologies without which To Be Continued would never have been possible, and what might be done to better understand, challenge, and redesign them in an age of artificial intelligence. Libraries, especially national libraries, are curators and stewards of culture formative in what counts as knowledge of our time and our place in the world. So it's very fitting that we are hosting a panel of researchers who all work on textual technologies from varied disciplinary perspectives, from varied periods, and varied places to explore the question of what reading and writing are becoming and what that might mean about who we will be. Please join me in welcoming Katherine and the panel. Thanks very much, Alison. I'd also like to acknowledge that we're meeting on the traditional
lands of the Ngunnawal and the Ngambri peoples and pay my respects to their Elders past and present. It's a privilege to be here at this wonderful place for writing and reading to consider the history of and continuing developments in textual technologies. That phrasing is a bit unusual, textual technologies, but writing and reading have always been technological, as perhaps librarians know better than anyone.
Because language is so central to what it means to be human, changes in these technologies often generate controversies. Socrates, the ancient Greek philosopher, dismissed writing in terms that are suggestive of some ways that ChatGPT is being discussed in critical commentary about so-called artificial intelligence applications. He argued that writing supplies not truth but only the semblance of truth and that those who write have the show of wisdom without the reality. When word processors were invented, commentators worried that these would lead to colourless, overwritten, and extravagantly qualified prose. Some authors concealed their use of these machines by having their manuscripts redone by a typewriter, and that was at a time when that word signified a profession as well as a thing. The same was true with word processor soon after.
Today, we're having another debate about changes in textual technologies focused on so-called "generative AI, which is an industry term, so there's the quotation marks. These applications use machine learning technologies and large amounts of scraped internet data to generate statistical models that predict sequences in words and pixels in order to simulate human cultural practises of reading, writing, and image making. This debate is in fact the background for three of our panellists being here with us in Canberra as visitors from South Africa and the United States of America for a planning meeting to prepare for a global humanities institute on generative... on Design Justice AI, whoops, that's the
industry term, no, Design Justice AI. That's going to be held in South Africa at the University of Pretoria in mid 2024. That two-week institute will bring together diverse researchers and community members to explore the perils and possibilities of AI, as we say in our grant proposal. The possibilities will be front and centre at the next session, which is to launch a new podcast made with fiction discovered in the National Library of Australia's Trove collection of historical Australian newspapers because that project was only possible with similar textual technologies to those we find in new AI systems such as machine learning and digitization of writing and reading. This session is here to balance that next one, to acknowledge that these
textual technologies also come with perils. With the use of large language models like ChatGPT, which is the most prominent one, the one we hear the most about, these perils are extensive. They include discriminatory results arising from algorithmic preconditions and training data used by these systems. There's now countless studies showing that they're likely to predict
that doctors will be men and that Muslims will be terrorists and so on, as you'd expect. There's also environmental harms. A recent study reported that ChatGPT requires 700,000 litres of fresh water in its training and that it drinks, as this study put it, half a litre of water for every 25 queries. There's also unconsented data surveillance and harvesting. Last but not least is the harm done to the unknown number of people, often from countries with low wages by global standards doing the human writing necessary to... often violent, degrading, and traumatising images and texts, doing that reading and writing necessary to pretend that these systems are either intelligent or artificial. So while we recognise these perils, and some
in the panel will talk more about them than others, I'm looking probably at you, Lauren, we're also going to think about how we can learn from the past, from other places, from other disciplinary backgrounds and research questions that are not often considered when we talk about these textual technologies, how they might be done better. So I'll do all the introductions upfront. Then we'll engage in a experiment in academic speed dating where I'll ruthlessly compel these international experts in textual technologies to condense their great knowledge into bursts of five-minute brilliance. Also, if you hold your questions till the end, we'll ask them all together. First up will be Kate Mitchell,
Professor of Literature and Director of the Research School of Humanities & the Arts at the Australian National University. Next will be Dr. Nomadloski Bokaba, Lecturer in African Languages and Literature at the University of Pretoria in South Africa. After that, we have Dr. Abiodun Modupe, Lecturer in Data Science and member of the Data Science for Social Impacts Group at the University of Pretoria. Then Lauren Goodlad, Distinguished Professor of English and Comparative Literature at Rutgers University and Chair of the Critical AI Initiative. Finally, Associate Professor Geoff Henchcliffe, Researcher in Digital Art and Design and Associate Dean of Education in the College of Arts and Social Sciences at the ANU. Please join me in welcoming our panellists. Thank you, Kath. I'm going to begin by taking us backward in time to ancient Greece,
probably Athens. Although on a cold winter Canberra day, maybe we want somewhere beachy here. I'll leave that up to you. It's a warm idyllic day, however, with the cerulean sky stretching above us, the faint scent of sea all about us as if the gods themselves had blessed this momentous occasion with their favour. In the middle of the bustling vibrant marketplace, an aged storyteller addresses the gathered crowd. Maybe he has a long white flowing beard cascading down his chest and deep set eyes that twinkle with the wisdom of centuries holding the promise of captivating tales. His voice rich and resonant carries across the throng, capturing every ear within its melodious grasp. As his words dance upon the air,
a vivid tapestry unfurls before the mind's eye. An epic battle springs to life: Zeus, the mighty ruler of Olympus, his voice thundering like a crack of lightning; Aries, the God of War charging forward with a ferocious battle cry; Athena, the goddess of wisdom brandishes her shimmering spear with grace and precision, her eyes ablaze with strategic brilliance. The Gods clash with titanic force, their powers colliding in a cataclysmic crescendo. Flames roar, thunder shakes the very Earth, while celestial beings weave through the tempest, their actions dictating the fate of both mortals and immortals alike. Now, as the old man's narrative reaches its zenith, a collective gasp sweeps through the crowd who, with their hearts racing united in a shared moment of wonder and awe, are transported to a world where gods walked among mortals. The ancient Greeks had a word for this kind of evocative wordsmithing that my imaginary aged philosopher has engaged with here, and that word was ekphrasis. Now ekphrasis means "to speak out
in full." It was a tool for the art of rhetoric. It was designed to bring something that was absent vividly before the mind's eye of the listener and, in doing so, produce a kind of affective, emotional response in its audience. Its purpose, in short, was to use words to paint a picture. Its purpose was to position the listener as witness to absent events, in this case of mythological battle, to participate through the speaker's powerful rhetoric in events that occurred elsewhere and long ago. The picture exists in the mind's eye. It's an
effect of both the words, the speaker's words, and the listener's imagination. Now, over the centuries, the concept of ekphrasis, both in theory and in practise, both narrows in its definition and expands or becomes more expansive in its use to think through various questions relating to word and image. Ekphrasis has transformed with various shifts in technology of both word and image in three key ways, in lots of ways, but I'm going to identify three today. As the written word became more readily available with the invention and widespread take up of the printing press, ekphrasis shifts from being a rhetorical strategy, something associated with the spoken word, to the written word. As it does this increasingly, especially after the Renaissance, becomes more and
more associated with the depiction in literature with literary poetic forms. The second is that as transportation technologies and shifting trade practises made travel more prevalent, ekphrasis took on a new life in travellers bringing back stories of the artwork they encountered as they travelled to Europe and through Asia. So ekphrasis becomes increasingly associated not with creating any sort of form of visual image in its listener or increasingly its reader, but referring specifically to an artwork that exists in the real world. So ekphrasis becomes the description of a work of visual art in verbal language.
Why think about this in relation to our contemporary moment? Through the centuries, ekphrasis is associated with thinking through various aesthetic sociopolitical kind of forms. But by our moment in the 21st century, its definition has expanded because of the expansive idea of what constitutes art with the introduction of digital media, digital artworks that might include interaction between the artist and the person engaging with the media, whether we think of that as a reader or an end user or a viewer. I'm just going to throw out a few key ideas. We've arrived at a moment in which, if the desired artwork does not yet exist, it's readily... sorry, we've arrived at a moment where artwork is readily available to us. We could all pick up our phones right now and
pick out any image that I threw out. We could find the image. So in the late 20th century, it was predicted that ekphrasis would actually cease to be important, cease to be needed. We can all see an artwork anytime we want to. But what we've found into the 21st century is ekphrasis has actually increased even as the ready proliferation of artwork and art images has increased and even as our relationship to art has shifted. Indeed, we've arrived at a
moment in which if the desired artwork does not yet exist, we can prompt AI to create it for us using words or prompts. Words make the image, and the image represents the words. It's a kind of reverse ekphrasis. So to pick up some of the issues that Kath mentioned in her introduction, the way that ekphrasis has been conceptualised over many centuries enables us to think through issues like, what is the effective power of word and image? What's its power in both the person who creates it and over the person who uses that image? The automated labelling of images, as Kath suggested, has highlighted for us the problem with centuries of practises of representation that produce biases in the way that we think about images. I'm going to close with one example that you might be familiar with, which is the painting the Girl with the Pearl Earring that became the book and became the film, the Girl with the Pearl Earring. What that ekphrasis does as an extended piece of
ekphrasis, it turns an artwork, it adapts it into a novel which is then adapted into a film. What the ekphrasis treatment of that does is highlight for us what happens at each stage of that introduction or that transformation from artwork into written form and back into a visual form. What the novel does is highlight the way the Girl with the Pearl Earring, if you know the book, it imagines that she's the servant of the artist. She's compelled to pose for this image. It takes
that sort of striking look that she has back at us, the viewer, and imagines that this is the way that this female servant asserted some sense of agency back when the image was produced. One of the ways I want to suggest that we might use ekphrasis as we move forward in this stage of generative AI that's being used to create words, words being used to create images, is to think about that intermedial adaptation from word to image to computer code, so we have a new instance of adaptation, a new adaptational intermedial moment there of which we're not quite in control perhaps. Certainly if you're a user like me, I don't really know what's happening in that algorithmic moment behind my prompt.
But being able to think about and separate out those different layers of how it phrases is working at each of those moments I think is one of the ways that we can unpick some of those inherent power imbalances, the new questions about what aesthetics means, what it means to create art, what it means to create literature in this new age. Thank you. Good afternoon, ladies and gentlemen. I come from South Africa, and I come from a country where textual technology is not an easy thing because we have a lot of languages which are official as given by the Constitution. A case study from South Africa on languages will indicate that the Constitution, which is one of the main document in our government in South Africa, declares 11 official languages, and recently, they've added the 12th language, which is South African Sign Language. *The languages that are regarded as official languages in South Africa will then be, this is Zulu, which is my language. It's a Xhosa, isiNdebele,
siSwati. If you check in that group, that four group, we call it Nguni group because all of them sort of start with, if you would've listened, you'll hear me saying isiZulu, isiNdebele, siSwati, isiXhosa. So those are like the isi-formative languages. If you know one of them, obviously you'll understand the rest. Then the next group is the Sotho group, where under we find Sesotho, Sepedi, and Sotho-Tswana. Those are non-isi languages. The other criteria is that the isi languages, we write them conjunctively, and the si languages, we write them disjunctively. In the conjunctive languages,
for I am working, I'll have [foreign language 00:21:19], which is one word. For the disjunctive languages, that is the Sotho group, I will have [foreign language 00:21:28]. So I separate the words that I'm writing. Already, I've calculated seven. Then we have two indigenous language or African languages
that are standalone, which is Xitsonga and Xivenda in addition to the languages. Why are they standalone is because when we check their linguistic formation, there's no relation to the isi group and the Sotho group. Hence, we say they are standalone. Then on top of that, then we add English and Afrikaans. Therefore, if we have to do textual technology, the Constitution requires that we treat all these languages equally and with equity, those two words. There's no discrimination against any other language. Any person who
is a South African, if I want to study in isiZulu, therefore the Constitution give me the right to study in isiZulu. Then the other thing is that for our government to try and cater the equity, equality, non-discriminatory effect in these languages, it then created a full department which contains a language practitioner for each language. You'll find that you've got nine language practitioners who are translators.
Then you'll have other nine language practitioners who are terminographers. Then you'll have nine practitioners who are human language technology, which is of course the part of the department that deals with textual technologies. We also have in that department, that department is called the Department of Sports, Arts, Culture, and Recreation. Therefore, the languages will fall under culture. Hence, they are in the department. In the department, we have a section that deals with language planning and policy of government. Meaning,
for example, like I said, South African Sign Language has been added as a 12th language, if we needed to do a law that the president must sign that sign language is now the 12th language, then each language practitioner will translate that document from English to Afrikaans into all these indigenous languages for the president to sign. So that if I come and say, "I want to understand what sign language is all about in Zulu," then I'll be able to get that document. Therefore, in South Africa, textual technologies will be highly assisted by the AI, but there are a lot of complications, and every person wants his or her language to be at the top of the ladder. Hopefully, we'll be able to do it right as South Africans. Thank you.
Good afternoon, everyone. I'm also from the University of Pretoria. If I have to talk about textual technology, I think I'll start by saying we all understand the world of technology or where we head the world of technology. Many of us, we thinks about our computer, your smartphones, and some other thing. But one of the things we don't actually always have at the back of our mind starts with, for example, your alphabet, the books that you read, the newspaper, and all those kind of things, or the television that you watch to be able to listen to news, and also the printing one, for example. So each of this examples that I give, they're basically technologies, if you understand it from that [inaudible 00:26:10]. They are basically technology innovation that actually has
changed dramatically in how we actually understand some certain things and how we perceive it and how also we can able to reason along with it. So this, we're now talking about textual technology, these are examples of, from here, how do we communicate? When you communicate perhaps, how do you also think along that particular communication that you are reaching out to other people? So for us, the questions in our research lab is, how do we now use all these language model collectively to be able to find a semantic representations of all these language that she has mentioned? Because we have a society that is very diverse in terms of how do we understand, in terms of how do we communicate, in terms of how do we reason together. Because I'm an African, and as an African you are talking about almost 250 languages. For example, I originate from Nigeria to settle down in South Africa. In Nigeria alone, even the government is saying
we have three languages, which is Hausa, Yoruba, and Igbo. In Yoruba alone, we have nothing less... I can count about 150 languages. So the question we should ask ourself from this textual technology and from what we know originally from the way we are brought up to be able to read... Because I have a three-year-old boy. If the mother wants for him to go to sleep, the mother needs to prepare him as early as possible, like 7:00 p.m., she will read with him, you understand what I'm saying, for him to be able to sleep. So the question is if I want my child now to understand my language, which is necessary for him to be able to speak my language, because if he doesn't, then my culture has already been diminished. So how will I be able to get a
book out there and to show that is in my language in order for me to read to him to understand? So if you look at it in the whole heuristic world, this language model that we have today, there are a lot of bias inside of them. So the question we are trying to look at our research lab, how do we now develop...? I mean use this technology test to be able to develop and share most of our texts, our contemporary text or textbooks or novel as well as so that we can able to understand the writing? Because I can write this way, but how do you conceive that particular writing? How do I be able to make that writer more explainable to you that whatever bias that you are concerned, I'm concerned, it's already been put in even while I'm getting this particular data? So that's just the most interesting things to us in our research lab. Thank you so much for listening. I think that's just what I wanted to share. I want to thank you all so much, and thanks especially to Kath Bode for inviting me to take part in this panel to talk about Critical AI literacy, a topic of increasing salience to all of us, I think. Critical AI at Rutgers, which I chair, is among other things the home base for Critical AI, the name of a new interdisciplinary journal published by Duke University Press.
You've heard earlier about the Design Justice Network's impact on us. Their principles and their focus on community centeredness have been a major influence on Critical AI from the start. It's the inspiration for the Design Justice AI Global Humanities Institute that will be hosted at the University of Pretoria next summer. But when it comes to AI in particular, building the foundations for a Design Justice approach will depend on the spreading of Critical AI research and literacies. What do I mean by that? Well, as we know, educators had barely recovered from the pandemic and from reintroducing dizzied students to in-person learning when ChatGPT came onto the scene in November '22. We're still in the midst of a well-funded hype cycle over this new technology with uncertain benefits and many known costs.
It should be obvious that there's no hurry to rush into teaching wholly new and untested methods. Here, I'm specifically thinking not about teaching data scientists, not about teaching computer science, although increasingly we learn that many if not most computer science professors do not want their students using chatbots in introductory courses for obvious reasons. They want their students to learn the fundamentals before they farm out their thinking to a digital tool. The same is true of writing. It seems like it should be obvious that there's plenty of time, humans have been writing for about 6,000 years, for us to not rush to changing the way that we teach the subject simply because a few powerful men who know absolutely nothing about education are now in an arms race with each other for monopolising the world's data resources. To my mind, worries of students cheating are way overblown. There are all sorts of ways
to engage students in strong writing. We need to be able to catch our breath, talk to each other, and talk to our students. That's part of what I mean about Critical AI literacies. In the fullness of time, we and our students may find lots of good reasons for using some chatbots in particular ways for performing the tasks at which they excel. And they do excel at many tasks. But let's be clear, college writing intended to teach research, communicative articulacy, and critical thinking is not one of those tasks. As many know, generative AI is a marketing term deliberately chosen to stand in for massive statistical models trained on unconsented data which generate human-like text through what Emily Bender, Timnit Gebru, and colleagues have called stochastic parroting, but which I prefer to call probabilistic mimicry. I hope to be able to make that clear what I mean by that. Despite misleading use
of anthropomorphizing terms such as neural network, this is not how your brain works. So when you hear about AI, you should not be thinking about this, much though I love this scene from Blade Runner and I bet you do, too, or even this, but rather of a vastly scaled up and multi-dimensional version of this. Please keep that image of a bell curve in mind. You may have already heard a lot about the harms of AI, including from Kath. The data
in these models train on and over-represent the discourse of white North American English speakers, often men of particular classes, while under-representing everyone else, including the roughly 30% of people on the planet who have never been on the internet. Because some of the best texts on the internet are behind paywalls, it also over-represents scrapable social media sites such as Reddit. Large language models are, therefore, larded with toxicity and conspiracy theories. They don't just reproduce human biases. They amplify the biases of particular demographics. Because they're completely untransparent, we don't know what's in them and can only guess at the harmful or malicious purposes to which they're already being put. You may not know this, but with social media sites, even though we don't know a huge amount about them, they are required to report malicious uses, and this has not yet happened for these very proprietary systems.
The preponderance of unvetted garbage-in means that, in ways that investigative reporters are just beginning to uncover, sustaining the illusion of artificial intelligence requires armies of low-paid human workers, millions of people with the potential to become billions, according to Google's recent paper. These are not great jobs. In fact, recently there's reporting that a lot of the people who are getting these jobs are farming them out to ChatGPT. I'll say more about that in a moment. The most traumatising labour, as Kath mentioned, is outsourced to workers in the global South. In addition, chatbots are environmentally irresponsible in every way, and they subject us to nonstop surveillance by plutocrats who have been lusting after the educational domain for decades. As a result, AI exacerbates and threatens to entrench political and economic inequality not seen since the last Gilded Age.
In the meantime, the pressure to teach our students tasks like prompt engineering, because this could be the job of the future, is an unexamined fantasy if you've heard that. Why? Because tech companies will gradually incorporate the most useful prompts into their models, just as for decades they've helped themselves to our personal data. While proprietary AI systems have scraped centuries of print culture in the public domain, a historic legacy that was digitised in the effort to create an accessible commons, which has now been put behind paywalls to sell to people like our students while subjecting them to endless surveillance marketed as effortless writing. This is from the Substack of a business professor at the University of Pennsylvania who lots of people read, "Setting time on fire and the temptation of the button," the button he's referring to is the button that Google is planning to put into Google Docs which will be something that you press and the button will say, "Help me write," and you'll press it, and it will start to shoot out a bunch of paragraphs. Here Mollick is saying, "We used to consider writing an indication of time and effort
spent on a task. That isn't true anymore." But I think that this is going too far. It isn't writing if it's not something that has time and effort because who's going to read it. It's probabilistic mimicry of writing supplemented by an exploited underclass of human labour through resource intensive conditions that benefit a tiny elite at the expense of everybody else. If you don't believe me about the limitations of this so-called writing, look what happens when you train language models on text generated by other language models.
The models collapse. See how the bell curve gets even narrower because that's how they work. They're predicting the most plausible answer. If you continue to go for the most plausible answer, you lose the tail at the edge of the bell curve, and then it stops working altogether. Think about that and about how that mimicry normativizes thought to the point of imminent model collapse. The next time someone suggests that maybe students, your children,
or grandchildren should learn writing by improving the first draught that comes from a chatbot or brainstorming with it. Now, you might readily see that the outputs that come from that draught are impoverished because you've been cultivating the necessary skills to do that all your life. But most students aren't ready to criticise bad writing any more than they would bad driving that might seem to them like good driving. They still need to learn how to drive. I know I did in college and even in grad school, and most of them do, too. So I'm suggesting to you today that there are a lot of things that we should be thinking about, educating ourselves about, and that there's no hurry whatsoever for us to adopt chatbots.
I have ideas for how we can help students to learn about chatbots and how we can teach ourselves about them, but that will have to wait for question and answer or another time. Thank you. Hello, folks. I'm going to offer a few thoughts as a designer of textual technologies. I'll start by reiterating an idea that we've heard already quite a bit, which is that design of a text is a socio-technical process, a practise in which practise informs technology and the practise itself is informed by the tech... practise informs technology, technology informs practise.
So with the design of text, we see the practises and forms evolving in alignment with their technology. In the design of textual works, volumes, we can plot it from the Gutenberg press through to industrialised forms of printing and then with the advent of computers, digitization, and more recently with the web and screen-based media. So we see this very neat and tidy technical evolution that I'm putting to you. Through all that, we see design practise and the forms we produce keeping a-step. And there's a dialogue there. One isn't necessarily always in control of the other. The technologies evolve,
we cop them, and we shape them through our practises and the forms we produce. So with the advent of AI and machine learning, I could posit that it'll just follow suit, that we'll just adopt this new technology. Sure, it will disrupt our design industry, it'll disrupt our practises, but over time, we'll co-opt that technology, and we will shape it through those practises. We'll work out what it can do, and we'll adapt and adapt it. Now, that's a very tidy and probably naïve and very optimistic view of these technologies. I don't want to suggest that anything here is inevitable, that AI is inevitable, or that our happy relationship with it is inevitable. As we've heard, I think there's probably more concern that it'll be anything but.
In terms of dangers, I think design is particularly vulnerable because it's highly programmatic. If we look at how we work with text and how we design text, we've spent a hundred years working out robust, systematic rules for how we present text so that it's accessible, legible, and it also means it's very easy to encode it in a machine. So, yes, design is incredibly vulnerable, but there may be opportunities, too. So I'm going to just think naïvely and positively for a second. I think one that's probably the most significant for designers of text is thinking beyond the types of text to the qualities of text. So as designers of text, we create templates for how we present
text, and it's based on what type of text is. This is a heading, this is a subheading, this is a body text, a caption, and we make rules and use those classifications in how we present it. Through things like AI, we'll have access to all sorts of other information about a text. What's the tone of the text? We've got a heading. But is it angry? Is it positive? Is it negative?
How is it coloured? And what can we do with that information as designers? That's quite a challenge. So we go from working just with the types of text to all sorts of other very subtle nuances of that text. A challenge, but also an opportunity. Another opportunity comes with access to all that content, that generative content. Now, this is an interesting one because it means as
a designer I can creatively explore ideas for how I want to present information, but I have no reliance on an actual human author. I can ask the machine to generate text for me. It's like the high-fidelity Lorem ipsum. We use placeholder text now. I can have pretty high-fidelity placeholder text. I can just use the machine to generate the content and the media for my designs. What that does is it potentially inverts a very well-established power structure. As designers,
we work in service to a text. That's our role is to represent that text, do it benevolently, make it accessible. What I'm suggesting here is that it inverts. I'm going to ask the machine to generate the text from my design. I don't like that text. Give me another one that's more suited to my design. Give me a media that's more suited to my design. So it's a fairly radical inversion.
I should apologise to all the authors and say, "Welcome to our world." Before AI, many processes of design have been ingested by the computer where they challenge the role of the human. So it's a relationship we're already pretty familiar with. What I'm proposing here is not a radical proposition. It's something that has
played out over the 20th century, and that is that kind of relationship between designers and their technology evolving in tandem, shaping each other. So it's a pretty optimistic view. I think more pessimistically, I don't think the barriers will be technological. I think a lot of the barriers to that kind of co-opt and vision will be cultural. It's whether we will have space as designers and artists to provide that disruption. I think some of the biggest impediment is the audience, that we ourselves are quite intolerant to that kind of difficulty that artists and designers will have to show us to the critical and creative, that turn that they will need to show us. In closing, I guess what I'm hoping is that, in seeing our culture endlessly imitated and regurgitated by machines like ChatGPT, that it will motivate us to come up with new forms of creative and cultural resistance, effectively new waves of punk.
Hello. Wow. That was quite a wild ride from the marketplaces of ancient Greece all the way to Silicon Valley, a stop by South Africa and the African continent and its linguistic diversity, and then ending with punk. I was not expecting that one. My brilliant experts have allowed us to have some time for questions. So if there's anyone in the audience that wants to ask a question, let me invite you to ask it now. The lights come up. You can all be seen.
Anyone? Okay, I will take... Oh, there we've got one from Paul Eggert up the back there. Thanks. Thanks very much. Oh, there's a microphone. Oh, right. Thank you very much to all the speakers. It's been a stimulating afternoon. This is a question that perhaps could go to the last two speakers. An AI generated text, we can ask the question of it, what does it mean? That is to say, we can put it to the text. We can think about what that text might mean. But I guess we can't ask the question,
what did somebody mean by it? I guess this raises the question of communication, what the writer may have been intending to communicate. This is something that's been pretty important to us in the history of our culture. Have you any reflections on its fate in the future? There's a microphone beside you. Can you hear me all right? Yeah, that's exactly right. The paper that I refer to as stochastic
parrots makes that very point that language models can't possibly express any form of intentionality in their writing and that the meaning that is implicit in the writing is there because the language embeds it and the person who reads it can interpret I so that something that used to involve a specific writer's... For me, the word intentionality is always a little bit difficult because I'm of the school that says writers don't always even really know what their intention is. But in any case, a person communicated something and did so with some degree of intention and another person read it, heard it, tried to make sense of it. Now you have something in between that is a statistical model. That's what Bender and colleagues call stochastic parroting, and I use the term probabilistic mimicry, because stochastic is an unfamiliar word, probabilistic is one you can all understand. Parrots are actually very intelligent animals that are smarter than
chatbots, and therefore, mimicry is a much better word to use. But you're completely right. Hello? Hello? Can I respond? I wouldn't disagree with anything that you just said there. I guess the most amazing thing that I'm reflecting on is just how good they are because they are just this brute force mimicry. So I think it's more revelatory about what we think we're doing, this amazing act of writing, and it's like, aren't the machines just revealing just how formulaic it is? Because they're able to do a very, very good impersonation. I don't know.
I think that we should take this moment to also reflect on our own practises and say, maybe we're not being as original as we really thought we were. How much of what we do is formulaic, is just based on cultural convention that we repeat and repeat and repeat again. In design, we constantly search out those formulas, and often we celebrate them. We actually point to them and encode them and share them. I think there's all sorts of that stuff obviously happening in all sorts of forms of culture to a degree that we are just not aware. The machines are showing it to us and saying, "It's not that original." There you go. I'm just going to put that one out for you.
I'm waiting for other questions, but I'll just say, Geoff, you mentioned that these machines are very good, and you're surprised by how good they are. We've been doing some experiments during our meeting that we've been having asking some of these generative AI models to do things in the English language and then to do them in non-English languages and discovering just how bad they can be and how flawed they can be. So I guess what I would ask is to the South African panellists on our panel, how do you feel listening to someone say, "Well, these are really great technologies"? Do you feel encouraged that maybe you will also be able to join in with them, or is there a feeling of discouragement? I am half/half to say the truth because even before the modern tools that we have now, the languages that we have in South Africa, especially African languages, haven't developed that fast even if we had translation tools which were our own that developed in South Africa for making sure that the languages that we have are moving faster, like English and Afrikaans which has moved faster than all the African languages. But the other half is that I feel because of the [inaudible 00:57:13], the way the world is moving faster, the new tools that we have now will maybe push these languages a little faster because of what is [inaudible 00:57:25] now. For me, I think the very first things that I think we need to do is to first and foremost just pause for a while and try and do an assessment of this particular language. For example, if I look at it from my own language, if I say a wife, which is [foreign language 00:57:52], because in my language, there's all of these... I mean from literature also perspective, taking morphology
and all those kind of structure that you need to, then I put it into Google Translations and the translation is giving me something else. It's giving me [foreign language 00:58:14], I mean my chest, which in my language, if you look at that, putting it in my language, it's the same word, those three word. But when I put it into Google Translation, it's actually giving me wife. Do you understand what I'm saying? So if you look at which context, is it really giving me that translation? So the question now is that we just need to... For me, we are of the opinion that, you know what, let us pause. Let us take all these language
model and pull them through because we're sure of the data that we are getting, that this data is actually coming from the people that speak the language, the people that write that particular text in their language. Can we now push it through all this language model and see what is the output of this language? By that way we can able to understand the insightfulness of that particular model or those representation before it can now go into production. I think for me, that's just the most important thing that we need to do, and we are trying to do in our research lab. Any questions? Can I have a...? Oh, we have one up there from... I sound like I've seated the audience with people I know, but I just happen to know the people asking the questions. One from Baden Pailthorpe up the back.
Hello. Thank you. I guess this is a question for everyone in response to a couple of things that the panel has mentioned, this idea of pausing or there's no rush to use chatbots. I guess just as I was listening to you speaking, I got an email, as I'm sure other people did, from the ARC saying there's a new policy that's just launched because some assessors were using ChatGPT to write grant assessments just recently for the discovery project round that's getting assessed at the moment. Also, the New South Wales Education Department is just talking about overturning the ban on ChatGPT for next year. I guess the question is the horse has kind of bolted to some extent, so what are some strategies perhaps to pause, as you suggest? I've been thinking about the same thing, Baden, listening to others.
I wonder whether... I mean, the horse has bolted, as we know it. If you're in universities, from our students, you know it, perhaps in our own practises sometimes. I think the thing for me that's quite interesting to think about is understanding the nature of what AI can do and what it can't at present, what these chatbots can do, thinking through the difference. I think one of the ways to look at it is what it's
showing us. Geoff said that it's excellent, and it is excellent at some writing, really. It's less excellent, I think, at the kind of higher-order writing, synthesising things, analysing things. It's that idea of parroting back, isn't it, or ventriloquizing what it's kind of trolled about and found on the internet, but it's not a perfect mimicry, or it identifies that gap perhaps between mimicry. One of the interesting things when I was just trying to play around with ChatGPT, I asked it to find incidents of ekphrasis, so verbal descriptions of paintings, in a novel by A.S. Byatt called Possession, which itself actually ventriloquizes Victorian literature. The author, Byatt, writes these short stories and poems in the style of Victorian poems and so on, and it's embedded in it. It was sort of identifying
passages of ekphrasis. I was like, "Oh, that's great." This is a novel I've written on quite a lot. I got to a couple and I thought, "I don't remember that in the book." Then it took another... not long because I know it well to think, "Yeah, that's not in the book." There were tells. You could tell that it wasn't actually Byatt if
you're familiar enough with Byatt's language. I think the thing that's interesting to me to think about is, because we probably aren't going to get a pause, how do we... maybe Lauren, you have thoughts on this in terms of educating in Critical AI, understanding what it is that the chatbot can do and what it is we're actually reading, which isn't a kind of straightforward English or other language. It's language that's kind of filtered through a statistical probability and kind of transposed back into language for us. I think there's ways to think more carefully about what that sort of journey through numbers, words into numbers back into words or images might mean or do or not do. Thank you. What Kate just referred to is a problem that through another... The industry,
the field of AI research loves anthropomorphizing terms. So it's called hallucination when a language model just makes stuff up. It does that because it has a very good sense of language patterns. So if I ask it to generate a bio for me, I'm likely to get something that gets about half of the information and half of it will be completely made up, awards that I never won, books I didn't write but somebody else did. But it will look
exactly like an academic bio because that's what it completely has grasped is the pattern of an academic bio. And it can only give you something that is plausible. It has no sense of true, not true, which is why so much human reinforcement is necessary. Never forget that, that army of human reinforcers that is necessary to give it even the level of accuracy that it has right now. As far as teaching with it, that fluency that you can get on something like, "Why is it important to have free speech in a democracy?" and in two seconds you'll get about four short paragraphs that will say this and that with no attribution whatsoever. The word plagiarism is perhaps misapplied, but suffice to say that it comes from somewhere, it's been synthesised, and it's unattributed. So if we were to teach our students, ourselves that it was fine to do that, we would in essence be saying it doesn't matter who said what, and it doesn't matter if it's right or wrong. I think really nobody wants to do any of that.
That higher-order level of thinking that Kate referred to, it cannot substantiate things well with evidence. Basically, at its best, on its very, very best day, it is giving you something like what is on Wikipedia about a given topic. It also can write very funny poems on improbable topics. That's another fun use. But as far as a source of information, it can give you basically on a good day, something like the Wikipedia entry, at which point you might say, "Why not just look on Wikipedia?" which will give you accurate footnotes and has been crowdsourced by thousands and thousands of people who have shared their knowledge, and you'll use less energy and water, which is what I tell my students.
I have zero problems so far with my students. I tell them I completely trust them, that they're not going to use chatbots in the writing of their papers. I know, I tell them that you have used this. You might even, for other classes, be asked to use it if it's, say, like a data science class where there is a technical purpose behind using it. I explained to them that they're going to be doing some good work with the bots so that they can say to, say, a potential employer, "I have used ChatGPT." What we do is we do probing experiments. What we do in that case is we probe the models to show the biases and the inaccuracies that are in them, which is a much better usage. It makes the student into a researcher rather than a consumer.
I also tell them that we'll be doing search experiments. So I have my students compare what they get from three different search engines and what they get from a database to what they get from, say, ChatGPT or Bing Chat. So in this way I kind of feel like we have the best of both worlds because the students are learning about the technology, they're seeing for themselves what its limitations are, but they can still go on a job interview and say, "Yes, my professor taught me all about that and made me a researcher." I think that's what we all should be doing. Well, I was going to cut you off Geoff, but go on, have the last word. Similarly in the educational space, I think we're talking with great caution, and I think people are observing those boundaries. But I'm struck by,
I think there's almost zero chance we'll slow it down, but not because I'm advocating it. Again, please don't get the wrong impression that I love this stuff, and I'm all on. I'm not at all. I'm just struck by people in the real world who are just using this stuff every day. They are also aware of it's not perfect, but it can do some things really blooming well, and they're into it. They don't just ask for the whole thing in one. They know how to coach it to the solution. So I'm just struck by how many people I encounter in day-to-day life who are
just using it routinely now to help them with these tasks, to generate the text, and it just accelerates what they need to do. I think so it'll be the weight of that kind of use, sadly, that will make it much harder to slow it down because it won't be from academics saying whether we should or shouldn't. It'll be the general public who are saying, "We love this stuff." It's built into so much... We're really aware of it now. One last thing. Note though, of just the image stuff. The image stuff has been deployed for ages. It's in Photoshop, but we seem to care less about that. The text really makes it plain to us what is going on here.
But it's the same thing. It's just like, "Oh, look, the bottom half of this image is missing, can you just fill it in?" It's like, "Yep, there you go. There's your image back." We're good with that. No one's talking about turning that off. Let's all go. It's interesting. I wanted to say lucky for us because our languages are still being grown or being developed so the data that is supposed to be in ChatGPT, it's not there. Therefore, we are still safe.
Well, on that paradoxical note, we could talk for a very long time about these new textual technologies. Thank you for joining us for this. There's lots we don't know about them. One thing I do know about people is that we shouldn't keep them from their food and drink for too long. So please join us in the foyer after this for some refreshments, and please come back for our launch of our podcast after this. Please join with me in thanking our panel.
2023-08-10