S2E4: Reading and AI

Summary Keywords

Generative AI, reading transformation, AI usage, large language models, digital literacy, AI governance, attention spans, deep reading, AI impact, copyright issues, human creativity, educational challenges, AI-generated content, ethical guidelines.

00:04 Host, Loh Chin Ee

Since the launch of the demo version of ChatGPT by OpenAI on 30 November 2022, the world has shifted into an era of ubiquitous AI usage. AI chatbots such as ChatGPT, Perplexity or Claude.ai can answer questions and write essays and code. Image generators such as DALLE and Midjourney can put together images at the command of a well-crafted prompt and Synthesia can create video from text in a matter of minutes. With such powerful tools, we have gained efficient ways to create new resources. Better still, there is no need for us to read lengthy articles or novels if we can get AI to generate short summaries for us. So, how has the nature of reading transformed in this age of AI? And what implications are there the way we write, publish and teach? 

In this episode of the How We Read podcast, we look at Generative AI and reading to ask if there is still a future/ for human reading and creativity.

Chapter 1: What’s AI doing to our Reading? 

Dr Chang Qizhong, lecturer at the English Language and Literature Academic Department at the National Institute of Education teaches academic writing and communication skills. He explains why Gen AI is such a game-changer in the language education space. 

01:39 Chang Qizhong

Gen AI Is Generative Artificial Intelligence. That's a system of artificial intelligence that focuses on generating content. So this content can be in text, images, audio and video. Large language models are a subset of generative AI, as the name suggests, there are models that are trained on large, oh, sorry, large data sets, right, large data sets of language, right, and They use certain things called transformers, and they use deep learning. They use neural networks to process and to contextualise and to generate information. 

02:22 Host, Loh Chin Ee

Prominent large language models or LLMs include Chat-GPT, LLaMA or Large Language Model Meta AI, Google Bard and more recently, DeepSeek. Qizhong sums up some of benefits and challenges of these technologies. 

02:40 Chang Qizhong

There's efficiency and speed. Yes, that could be seen as a good part of Gen AI, but if you think about it in the other sense of, for example, if I want to produce bots on social media to mislead people, right? I can also do that at scale. So, you know, having skill is, is it good, or is it bad? Objectively, if you look at it neutrally, it's good, but it really depends on the way you use it. So it allows more people to use it. It allows for people to reach audiences they have not been able to reach before. So for instance, if English is not my native language, I write in a different language, I can translate things very easily and the same way it goes. If I want to access something that is not in my native language, I can get it translated as well. So it's supposed to democratise the whole process and be as inclusive as possible. So I think that is, generally speaking, a proven

03:39 Host, Loh Chin Ee

Professor Simon Chesterman, David Marshall Professor in the Faculty of Law and Vice-Provost of Education Innovation at the National University of Singapore, is a Senior Director of AI Governance at AI Singapore. His most recent academic book is We, the Robots? Regulating Artificial Intelligence and the Limits of the Law. I ask him, how has AI changed the way we read? 

04:05 Simon Chesterman

So one of the challenges in the AI space is everything moves so quickly. And in fact, a kind of revelation for me as a lawyer, where the stock in trade is books and articles. In computer science, things are changing so quickly that often the cutting edge work is presented at conferences, almost live, and working papers often quite rough. Working Papers the main vehicle for developing cutting edge research. And so I think one challenge is the diversity of ideas, the diversity of disciplines involve means you've got to move outside your comfort zone, so I'm reasonably confident reading almost any legal text about any legal matter. But the more you get into computer science, the further you get from my expertise. The reason I've pursued and many others have as well, is that AI is going to be so important, you can't just leave it to the computer scientists, but nor should you just leave it to the lawyers, because the lawyers don't understand the computer science, what you need is a kind of collaborative endeavor. And that can be a challenge in the material, because it's not always accessible, and so you risk just skating across the surface.

05:07 Host, Loh Chin Ee

With so much information available out there, there is a tendency to skim and scan when reading information online. While skimming and scanning are useful skills for getting at key ideas, what happens if students only read summaries of articles rather than the originals? Is there room for deep and critical reading? 

05:28 Simon Chesterman

This is a huge challenge as an educator, because it's very easy for students to fake knowledge now, to get digests and not to do the hard work of going and reading. And I suppose the unknown is how important that kind of deep knowledge is going to be in the future. To pick a current example, there are really interesting tools out there. Just one that I found, striking notebook, LM, something released by Google. They can produce plausible podcasts that might sound as realistic as you and I talking right now on the basis of material that can digest in minutes. One interesting question is whether people will continue to spend the time that's necessary going into depth reading. And this is a more recent version of an older problem, which is just attention spans. There's some studies that basically suggest students aren't reading as much as they used to, and they find it hard to read as much as they used to. One reason I kind of force myself to read novels is to check that I still have the attention span to do that, whereas a lot of young people today will sort of skate across the surface. One of the challenges as an educator is how much preparation we can expect students to do, and whether they will do it, because that sort of breadth versus depth. If everyone's read the material, you can have a much more interesting conversation if no one's read the material. It's like a book club where no one's read the book, and you can sort of talk about what's on the cover, but you never really get into the meat of it. And the challenge for us as educators is working out well, what's realistic to expect of our students, and how do we incentivise them to do the work that we think is necessary without being so unrealistic that we leave all but a couple of them behind.

07:06 Host, Loh Chin Ee

Neil Humphreys, a writer, journalist and broadcaster, fears that our young people may not read deeply today. 

07:13 Neil Humphreys

There’s a recent article that went viral. You probably saw it that this American University professor said he couldn't believe that his bunch of undergrads had never read.

07:24 Host, Loh Chin Ee

The article is The Elite College Students who Can’t Read Books, published in the November 2024 issue of The Atlantic. The writer, Rosa Horowitz, interviewed 33 university professors in the US, many who bemoan that their students have difficulty in sustaining attention, are less able to persist with reading through a challenging text and have reduced vocabulary and understanding of language. They attribute this in part to the students arriving in college without having read a single book in their middle or high school, because it was not assigned to them.

07:59 Neil Humphreys

This is the first generation where this has happened, meaning they can get all the way to university without reading. They'll read paragraphs and excerpts of maybe Shakespeare or Chaucer or whatever Hemingway, whatever it is, but they don't read in the way that previous generations did. Every generation read, right whether it was your Enid Blyton or Roald Dahl or fantasies. You've got older, but every generation read. This is the first generation that just doesn't have to read. Do you see the difference? They can still do very well in their coursework and their exams, but they don't have to read anymore. They can read paragraphs. They can read excerpts. so, by the time you get to university, you've got these entire huge literary gaps in your knowledge that were taken for granted a generation ago. And it doesn't matter. That concerns me.  

08:51 Host, Loh Chin Ee

Chapter 2: Gen AI’s Impact on Publishing

When we think about learning, a human being and GenAI learns very differently. When I, as a human write a poem, I am relying on what I have read, what I know and my own life experiences to generate a new poem. When ChatGPT writes a poem, it is using statistics and the pool of data available to it to generate the poem. Qizhong explains that the quality of a LLM really depends on the quality of the input. 

09:39 Chan Qizhong

Basically, whatever we use Gen AI to produce, right now, the output can form the basis of new input for Gen AI to learn from itself. And in that sense, if I produce bad output, then AI will be learning from that bad output, bad in inverted commas, right? So when AI learns from bad output, then it produces even worse ones, so progressively it deteriorates. So given enough time, this could lead to model collapse, which is very gradual, maybe not even gradual, the deterioration of the input and then leading to progressively worse output.

10:15 Host, Loh Chin Ee

The need for new input is why the large tech companies are looking to find even more data for training its models. It is expected that the AI industry will soon reach a “data wall” when they run of out of quality data for training their models. 

The issue of the kinds of data that AI can ingest is being put to the test. On 27 December 2023, The New York Times filed a lawsuit against OpenAI for using copyrighted works from news organisations for machine training, without paying or consent from the news organisations. Simon explains the issue. 

10:53 Simon Chesterman

So this is where there's litigation and fights and real dilemmas, and I would think of it in kind of two categories. There's the sort of short-term question of the economic viability and the threat to publishing, and then the long term question of what happens to information. In terms of the shorter term, there are lawsuits basically over the way in which large language models have been trained on data sets that can be thought of as falling into three buckets. Basically, these data sets, what the training data of something like ChatGPT, it's trained on the entire internet, among other things. And you can dye that into three buckets. There's the publicly available material, so the works of Shakespeare, the Bible, all the stuff that's publicly available that's fair game, where the author's been dead more than 70 years. Public Domain material, no question, you can use that for whatever purpose you want. The other extreme, pirated material. There are things like Books3 Dataset, Peter Shopper and others have highlighted, he's from NUS press, and the Singapore Book Publishers Association highlighted, there's a lot of pirated material that should not be on the internet, and if you're accessing that, you're furthering the sort of harm associated with piracy. The middle ground is the trickiest, that's copyrighted material, where the large language model creators are arguing that this is fair use in the American term, and that's sort of the heart of intellectual property. Is intellectual property is designed to incentivise and reward human innovation, because we want people to create. We want them to be able to make a living by creating it, but you also want people to be able to use it. And so, where we end up is a kind of tension where if you write a book, you can have the right to make money out of that book, but we want other people to be able to read it, to quote from it, to be inspired by it. And so if I want to quote a paragraph from Harry Potter in one of my academic works, I can, but if I want to reproduce an entire novel, I can't. The question for large language models is, what they're doing. Is it more like me reading the Harry Potter books and then being inspired to write my own novel about a young wizard? Or is it more like me taking the Harry Potter books, cutting them up and rearranging them? And arguably, it's a little bit more like the latter, particularly because, and this is where the lawsuits come in, the production of these works will undermine the economic viability of future sales of that original work, but also of the medium more generally. And this is the essence of the New York Times suit against ChatGPT against open AI, where they're saying, simplifying somewhat open AI has a subscription to The New York Times, but through its outputs, it is reducing the likelihood that people will buy subscriptions to the New York Times because it's competing with that original work. So that's the kind of short term and then the longer term question is the quantity problem. This is the fact that AI is not only good at producing human like high quality content, but vast quantities. And so, there is a real risk that we will just become swamped in what's called AI slop, or even high quality AI stuff. One more thought is my favorite statistic in the AI space is that Amazon is one of the world's biggest publishers, mostly self-published books these days, and it had to introduce, about 18 months ago, a limit to how many books you can publish. And so the limit to the number of books you can publish through Amazon today is three every day because AI was producing hundreds and hundreds of books a day, at really low numbers, but there is this risk that you will have this sort of vast amount of material that will make it very hard to sustain a creative sector for anyone to make a living as an artist.

14:19 Host, Loh Chin Ee

At the heart of this matter are issues of human creativity and the value we place on the contribution of authors and artists. If we believe in the value of literary work, we must ensure that authors have the means to live while creating works that entertain, inspire and nourish.  

Chapter 3: Educating for Reading

If it’s true that our students’ attention spans are limited today, how are educators dealing with it? I ask Simon how much reading he’s assigning to his law students. 

14:52 Simon Chesterman

It depends on the content, but I think the days of assigning multiple books to a student every week, those days are probably gone. And and indeed, there's a pretty good argument that if you can get students to read a really complex text but spend like a lot of time with even just 10, 20 pages, that's better than nothing. But we don't really fix a limit at the university. It depends partly on the subject matter. Sometimes you're assigning multimedia, but certainly we do expect students to prepare otherwise when they come into class, if the only way you're transmitting information is orally, apart from anything else, it's just inefficient. Again, the challenge as an educator is, what is our job? It used to be there was this metaphor of sometimes attributed to Aristotle that students are empty vessels, and we the professors, fill them up with knowledge, or the teacher fills them up with knowledge. But now, of course, we all know that there's much more information on the internet than we could ever have in any one person's head. So we've been talking about skills and critical capacities, but that's something that AI is getting better and better at. But as we think about time in the classroom, if you don't at least come in with some knowledge, some information that you've already digested, then again, all you can really do is talk about surface matters, and it's very hard to go deep.

16:03 Host, Loh Chin Ee

Simon suggests that when it comes to reading some materials, the reading is the point. The materials might be difficult but it is the process of trying to make sense of the materials that counts. Habitual engagement with words and ideas trains the mind to go deep and think critically.

16:25 Simon Chesterman

It’s important that students and the wider public understand how this is being produced, that it's identifying the most probable next thought, next segment. It's creative in one sense, and a lot of human creativity actually isn't that creative? I mean, there's various studies that suggest there are either three or seven plots, and every novel is a version of that. But the concern I would have about using these tools is that it assumes that the work is the output as opposed to the input. Like the purpose of a university education is not that you produce x number of essays and write x number of exams. That's the output. The input is what's important, the ability to process information, analyze it, produce it, produce your own thoughts in a way that is interesting, compelling. That's the work. And this is a larger problem associated with AI that makes a lot of things easier, but sometimes the effort is the point. I mean, you can get the gist of a book by getting notebook LM to produce a podcast, and you can listen to it over 10 minutes or so, you'll get the gist, but you can get the gist by reading CliffNotes. There have always been shortcuts. The dilemma we're going to be confronting in the very near future is how important we regard that grunt work, that effort, that effortful learning, as part of an educational journey, and to what extent that will be important, not just at university, but in the workplace, where AI is going to be even more transformative than I think it will be at universities.

17:55 Host, Loh Chin Ee

Reading and writing are intimately connected. And good writers are often good readers, who read widely. Whether we can persuade our students to do the grunt work of reading is going to be a crucial aspect of teaching. 

When it comes to writing, Neil reminds us that human inspiration, emotional engagement and creativity is something that humans do much better than AI. 

18:24 Neil Humphreys

I'll give you a quick example that I often share. I saw two examples. I saw a model essay that was written in the tuition centre in the East Coast. This student went to this centre and produced this paper and got an A grade. That paper would stop me sending my daughter to that tuition centre, because it didn't look like it was written by a real person. It looked like it was written by AI before AI became what it is today. Number two, on Facebook, I saw this example, a tutor. I'm not picking on tutors. I've been. Tutor, I can see, of course, I can see the merits. But he said, this was the piece written by this 10-year-old boy when he handed it to me, this is the piece when I'd finished with him. You know what I'm going to say? I preferred the first one because the first one was a realistic, visceral experience of a 10-year-old boy. It was raw. It was a little bit inelegant. The language was simple, but it was real. It felt real to me. It got me in the gut. Second one was AI. It might as well have been written by AI. I could even join the dots. I could see the cogs turning. I could see where he said, “you need an adverb,” “there you need…” I could feel it. I could seal it. I felt nothing. I felt cold. I felt detached. If, see, I'm getting passionate. Now, if this is the way we are raising our children in Singapore, we have already lost, we have already lost the battle to AI, except the fact that all your creative endeavours will be done by Ang Mohs or Japanese or South Koreans, if you do not change now, if you do not allow young people to explore, to make mistakes, to fail, to try and find their own voices, to try to write their way, not the AI way, not the way that's going to get them an A plus plus plus plus, like , you know, it’s some restaurant bill, they’ve already lost, because AI will wipe them out like a tsunami.  

20:14 Host, Loh Chin Ee

As we already know, AI is not perfect. ChatGPT has been criticised for its tendency to “hallucinate”, a euphemism to explain that it generates misleading and false data. The quality of the product is only as good as the quality of the dataset it is trained on. Depending on the datasets that each version of GenAI is trained on, results may be biased or limited.  

To know when to use GenAI and to use it effectively, we need to be digitally literate. Digital literacy goes beyond technical skills and safe usage of online resources. It includes the ability to search for AND critically evaluate the information that we encounter. 

To combat uncritical use of GenAI, Qizhong brings it directly into his writing classes. He uses it as a tool to help students improve their writing process. He also encourages them to critically evaluate how AI is used.  

21:20 Chan Qizhong

I try to incorporate tasks to illustrate that AI generator output might not always be the best. It might dissuade them from being too uncritical in how they use AI, because if they have it in their heads that it always produces perfect and beautiful output, then they will use it very uncritically. I will pull up the program on the computer so it's projected on the screen, and I will type in some prompts. I will get the output, and then everyone can see for themselves that the output is not ideal, not something that they want to be producing in their own essays. So I'm trying to kill two birds with one stone. So firstly, the importance of the prompts, how you ask the question, and secondly, to convince them that AI output is not at a level where you can just throw anything at it and you get what you want. Sometimes we take human, written samples that are bad, and we try to evaluate what's bad about it. Then sometimes we take AI return samples, and then we evaluate what's bad about it. And then sometimes we get students to put their own writing into AI and have it evaluated or critiqued, and then we discuss whether the critique makes sense or not, because sometimes the critique doesn't really make sense. So those are things we do. Can the students tell whether something is written by a human or an AI? Not that well, but it also depends on what you get them to look at. So if you ask them to look at whether the piece of writing is being critical, whether the piece of writing uses reliable information, whether the piece of writing is well organized, depending on the kind of questions you ask, you will get better or worse responses.

 23:15 Host, Loh Chin Ee

Looking into the future, Simon reminds us that we are far from the death of reading in this age of Generative AI. 

23:23 Simon Chesterman

UNESCO did a big study on ethical guidelines for generative AI, and the one thing that they could agree on is to keep it out of the hands of children. The challenge really is getting people to be intentional about technology. And here the risk is just starting with families, there's so much sort of pressure on parents to monitor screen usage and so on and how children are using their devices. And that seems like kind of empowering families, but what it means is that the families are against some of the most powerful, wealthy corporations in the world with highly trained people whose job is to monetise attention. So it was a clearly unequal battle, and so that's why there are big debates about sort of, at least age restrictions, and then when you're looking at older people, I think it's hard to control what adults do. I think we've got to incentivise people. I suspect anyone who's listening to a podcast like this, we're not so worried about because they're interested in reading. But how do you make that attractive to a larger number of people without just sort of giving up and reducing culture down to something that's easily digestible in sort of 15 second videos on Tiktok. And I think that's going to be a dilemma, but we've grappled with this in the past. When movies came out, they said, Well, that'll be the death of theater. It wasn't, I mean, it did change theater when television came out, they said that would be the death of reading. It hasn't been quite, but I think it's going to continue to be a struggle.

24:44 Host, Loh Chin Ee

So, to conclude, there is still room for human creativity. And reading remains important in this age of GenAI. Whether we continue to read, and whether we will read differently in the next decade remains to be seen. But if history is anything to go by, there will still be room for reading. 

Thank you for listening to Episode 1 of the How We Read podcast episode where examine the role of Generative AI on reading, writing and publishing. Many thanks to my interviewees for their human input on this issue. If you would like to hear more, we have two bonus episodes this week. Neil Humphreys shares about writing and Prof Simon Chesterman shares about reading the law. 

This episode was written and hosted by me, Loh Chin Ee. It was produced by Kenn Delbridge of SPLiCED Studios. Administrative support was provided by Koh Yu Qun. The Singapore Book Council provided funding for Series 2 of the How We Read podcast. 

We’re available on all major podcast apps so do subscribe to be notified of new episodes and take a moment to give us a 5-star review.  

Swipe on the cover art to see show notes with links and references.  

For more information, please visit lohchinee.com