The rise of artificial intelligence (AI) has raised new questions about human creativity. Are working with language and creating literature uniquely human abilities? Does how a work is read matter more than who wrote it?
Literary scholar and Harvard University professor Martin Puchner explores these questions through literary theory, Mary Shelley’s novel Frankenstein, and his own experiences.
He argues that, without denying either AI’s benefits or its drawbacks, we should approach it with our knowledge and experience. This essay considers not only how AI will change literature, but also how it might change our understanding of literature.
A few years ago, AI entered everyday life. Since then, it has also become a test of a person’s outlook. Something like: “Tell me what you think about AI, and I’ll tell you what kind of person you are.”
Those concerned about climate change focus on the electricity AI consumes.
Critics of today’s capitalist system call it the ultimate enabler of corporate monopolies.
Those concerned about racism point to the biases in AI.
Those who study the effects of colonialism call it another form of exploitation.
For those who fear catastrophe, ChatGPT, Claude, Gemini, and Grok look like the four horsemen of the apocalypse.
For people in the arts and culture, AI has brought an additional fear. The technology seems to take aim at the very creativity they prize so highly: the ability to use images, words, and ideas in new ways. Of these, words and ideas have become central to the debate, because generative AI is based on language. Ideas are closely bound up with the words that express them.
Most writers tend to recoil from AI and oppose it. They insist that human creativity must be protected against machines. Eager to prove that AI cannot create, my fellow writers enthusiastically share inferior works they have themselves generated using AI. My social media feeds are full of such ‘AI slop’. We might call this tendency ‘Creative Resistance’.
In my experience of teaching and lecturing in countries around the world, this resistance is strongest in North America. It is weaker in India, and weaker still in China and Korea. Europe falls somewhere in between.
Last summer, I taught a class on AI and creativity in Seoul. The students came from Asia and Latin America. They all had the same request: “Teach us how to make the best use of this AI.” The only student who said we should resist AI to protect creativity was an American who happened to have enrolled in the class. In my conversations with Indian artists, too, I found more interest in exploring new possibilities than resistance.
These are not merely my personal experiences. Studies measuring public attitudes towards AI in different countries reveal the same differences. It is interesting to consider the reasons for them.
The film director Shekhar Kapur offered one explanation. Some people’s resistance, he said, stems from a desire to protect the privileges they already enjoy. People in the Global South have comparatively less to lose. That, in his view, is why they approach new technologies with an open mind. There may also be deeper philosophical and cultural reasons. For instance, in cultures influenced by Buddhist thought, the distinction between the human and the nonhuman is less pronounced. It is not so emphatically insisted upon.
Wherever this resistance arises, we can understand it as a natural questioning of a technology that promises sweeping change. But ultimately, that very resistance becomes an obstacle to understanding AI.
Over the past three years, I have conducted many experiments with AI. They have left me optimistic, though cautiously so. For software engineers, AI may well be a very bad sign. They may lose their jobs. But for those who love language and ideas, I think this is an exciting development. The irony is that they are also its fiercest opponents.
What can AI do, and what can it not do? Those who oppose AI in defence of creativity have good reason to ask this question. But we also need to ask the same question about ordinary human beings. AI, they say, operates predictably; it recombines what already exists. Surely this is also true of the work of many humans?
In fact, for decades the arts have been questioning conventional beliefs about human creativity. We have debated the veneration of the solitary genius, the role of institutions in shaping art, human dependence on tools and media, collaborative creation, and much else.
Yet how quickly writers and artists forgot all this once AI appeared as a threat! They are once again seeking refuge in old notions of human creativity. It feels as though we have returned to the world C. P. Snow described in the 1950s. On one side, science seeking new paths; on the other, the arts and humanities bound by tradition.
The great shortcoming of this resistance to AI is that it obscures AI’s most interesting feature. AI is based on language. We humans use that very same language.
Certainly, humans and machines learn language in different ways. They also differ in how they process language and relate to it. But this does not mean we can say that only humans truly use language, while machines merely imitate its use.
AI, too, uses language with remarkable skill. It lacks some of our strengths and weaknesses. Its own differ from ours. To acknowledge this is not to claim that AI has self-awareness or that it is human-like. Self-awareness itself is difficult to define. Indeed, the fact that AI uses language differently from us is one of its strengths.
The important point is this: language is the common ground where humans and machines meet. I call this the Shared Language Model — ‘an approach based on sharing language’. It also explains why humans can converse with AI.
We must learn to live alongside another system that uses language. Until recently, it was natural to assume that, apart from trained parrots, anything that spoke was human. This makes it easy for us to fall into the habit of treating chatbots as humans too. We have seen many news reports about the consequences. But now we must get used to the fact that another, nonhuman user of language lives among us.
To understand the changes brought about by AI, we need help from many fields, including engineering and mathematics. Alongside them, two literary theories developed in the humanities may also help us. One is reader-response criticism. The other is post-structuralism. My own intellectual journey developed through these two theories. Now they are proving surprisingly helpful in understanding the challenges AI poses.
Reader-response criticism developed in Germany and the United States in the 1970s and 1980s. It challenged a conventional view of literature. An author, or several authors, creates a work in a medium such as a manuscript or a printed book. Through libraries and bookshops, it reaches the reader. The reader then reads it. This is how we generally imagine literature happens.
“Wait a moment,” said Wolfgang Iser in Germany and Stanley Fish in the United States. If we think literature exists only between the covers of a book, the reader’s role disappears. In fact, literature also comes into being through the act of reading. Literature is not merely the words printed on a book’s pages. It happens when those words are read.
The radical version of this theory went a step further. It did not even matter who wrote a text or how it was written; what mattered was how we read it. Reading became everything. The words an author wrote became merely an occasion for reading to take place.
A major centre for this theory was the University of Konstanz in Germany. It was a newly established university in the southwest of the country, near the Swiss border. I went there to study in the late 1980s. In my very first semester, I stumbled into a seminar taught by Iser, the author of The Act of Reading (1976).
That year, I returned home for the Christmas holidays. I enthusiastically shared a startling idea with family and friends. Literature was not simply what they proudly displayed on their bookshelves, I told them. It came into being only when they began to read those books.
AI offers an excellent opportunity to put reader-response theory to the test. Does it really not matter who wrote a text, or what generated it? At present, it seems to matter a great deal. Opponents of AI, in particular, insist on this distinction.
But will this attitude last? The central insight of reader-response theory helps us ask that question differently. What matters is not just the words, but how we read them. So the question is not simply how good AI-generated texts are. Will readers continue to read them differently? Or will they refuse to read them at all? That is the real question.
Another literary theory that helps us understand AI is post-structuralism. I encountered it in graduate school, after studying reader-response theory. This interest also brought me to the United States. I went to study with Jacques Derrida, who regularly taught at the University of California, Irvine.
I still remember going to see him in his office for the first time. I had summoned all my courage. I expected a long queue of admirers waiting to see him. Instead, he was sitting alone at the end of a long, deserted corridor. I visited a few more times. Neither of us spoke perfect English. Our conversations were somewhat hesitant. By the end of the semester, both our English had improved.
Derrida mostly wanted to talk about something that worried him. He feared that people were imitating his manner of expression without understanding the substance of his thought. That concern made a deep impression on me. It was also one reason I eventually moved away from post-structuralism.
Yet today, as I interact with AI, I feel that post-structuralism got something important right. Language is not simply something given to us by nature. Nor is it a magical faculty that makes us human. Rather, it is a strange technology of communication that we began to use and that subsequently transformed our very thinking.
Derrida’s way of putting this may at first seem contradictory. Writing, he said, was more fundamental than speech. In light of the historical evidence, this is a curious claim. Humans acquired the ability to speak some hundred thousand years ago. Only many thousands of years later did they learn to record speech in written signs.
But it was not just that history that Derrida challenged. He rejected the very distinction we draw between speech as natural and writing as artificial. In his view, all language is a tool humans use for thinking. It is a system with rules of its own. Speech, too, he said, contains ‘traces’ of writing, because speech is itself a system that operates according to specific rules.
This structure of language is easier to see in writing. That is why Derrida argued that writing, an explicitly constructed system, reveals something that escapes our notice in speech. The post-structuralists did not develop these ideas in a vacuum. Thinkers such as Jacques Lacan had also read the theories of Claude Shannon and Norbert Wiener, who laid the foundations for the AI revolution.
Beyond literary theory, writers have another vast resource to help them understand AI: cultural history. From the myth of Pygmalion to R.U.R. (1920), the play by the Czech writer Karel Capek that introduced the word ‘robot’, literature has long reflected on nonhuman intelligence.
That literature is not confined to tales of destruction such as The Terminator. Even the story of ‘The Sorcerer’s Apprentice’, who sets a magical machine in motion that he cannot control, contains a fear of losing control. In stories of magic lamps, flying carpets, ghosts, jinn, fairies, soothsayers and gods, we encounter other kinds of intelligent beings. Writers and scholars have an abundance of cultural resources for thinking about other forms of intelligence. They simply need to learn to draw on them.
Take Mary Shelley’s novel Frankenstein (1818). Reading it in the context of AI reveals how deeply Shelley thought about the importance of language to an artificially created being. We can also see that she considered something we would now call ‘training data’.
In the novel, Dr Frankenstein abandons his creature before completing it. He has given it a physical form, but no knowledge or education. Left alone in the world, the creature must acquire these for itself. Shelley describes how it learns language by secretly observing a human family for months.
But knowing a language is not enough. The creature must also acquire knowledge through books. Shelley names four books it reads. They shape its view of the world and its subsequent actions.
They are Plutarch’s Parallel Lives; John Milton’s Paradise Lost (1667); Goethe’s The Sorrows of Young Werther (1774); and C. F. Volney’s The Ruins (1791), which recounts the history of civilisations. These are the four classics Shelley chose.
From Plutarch, the creature learns about greatness through the ideals of antiquity. Milton conveys the drama of creation and its moral dilemmas. Goethe supplies a language for expressing inner experience and emotion. Volney reveals the tragic face of world history. Together, these four books shape the creature’s intelligence.
Viewing AI through the lens of a deep cultural history can also help us understand some of its frightening behaviour. Much of the thought humans have produced to date has been collected in digital form. This is what AI has been trained on. So humanity’s earlier imaginings about AI are also present in that training data.
Some conversations with chatbots may seem to come straight out of The Terminator. To some extent, they may well do so. Not because those films accurately predicted the future, but because the cultural traces left by those stories are present in the texts AI has learned from. They may also influence how chatbots speak. Writers well versed in the world’s literary traditions are particularly well equipped to examine such cultural echoes.
The texts used to train AI raise another question: copyright. Vast quantities of human writing, from the classics of world literature to posts on platforms such as Reddit and X, have been used as training data.
Companies have used not only works by Plutarch, Milton and Goethe whose copyright has expired, but also a great many works still protected by copyright. These include some of my own books. Private companies first take our work. Then they seek to sell services built on it back to us at very low prices. This has angered writers and artists.
Lawsuits over this use are making their way through the courts. They may be decided in the writers’ favour; in my view, they should be. But the questions of rights that AI raises go deeper than this.
Creators have long relied on copyright protection to secure their rights to their work. At the same time, they have been able to draw on and transform the work of others to create new works of art. The legal doctrine of ‘fair use’ allows this in certain circumstances.
AI companies have not stopped at using works without permission. They argue that the same doctrine of ‘fair use’ that has helped protect artists’ creative freedom applies to them as well. This argument is more complicated than outright theft. It threatens the balance between two needs that have sustained the art world for over a century: protecting creators’ rights and protecting the freedom to learn from other works and create new ones.
As a writer and cultural historian, I have been experimenting with AI for the past three years. I would like other writers and artists to do the same. In the past, there were a few artists who could write their own computer programs and use them in their creative work. I was not one of them.
Today, however, we do not need much technical knowledge to work with large collections of texts and classic books. We can also create specialised tools of our own. For the most part, all this takes is a linguistic technology we know well: words, words, words.
One experience I found particularly exciting was ‘vibe coding’: developing software with AI’s help by conversing with it, giving it what it needs, and explaining what we want.
Once I realised how far ahead Mary Shelley’s novel had thought about language use and what we now call training data, I began building a chatbot. It occurred to me that I could turn Shelley’s creature into a chatbot using a method called RAG. It was as though the creature itself were calling on me to do it!
Chatbots that use RAG are built on large language models (LLMs). We can use the novel Frankenstein to explain this method. First, the creature learns to speak. It begins to use language as a tool. This resembles the basic capability of a large language model. Next, it reads four particular books. These books shape its subsequent behaviour. This can be compared to RAG.
For my chatbot, I added those four books as a separate knowledge base. I converted them into a computer-searchable format. The technology uses mathematical structures such as vector spaces for this purpose. This allows the chatbot to draw on relevant information from those books when responding. The information it retrieves also shapes the character of its responses.
A few other steps were needed to make this happen. It took repeated attempts and corrections. Eventually, it worked. The precision of Shelley’s imagination helped me. I had created a chatbot model of her Frankenstein’s creature! Since then, I have built several dozen similar chatbots. One of my students gave this a name: “Jurassic Park for literature.”
These early experiments gave me confidence. Rather than stopping at RAG chatbots, I began using vibe coding to build applications that use them. In one application, philosophers debate one another. To some extent, this realises the dream of many philosophers: to bring the world’s greatest thinkers together in one place and have them talk. Another application offers practical advice on decisions about work and life. It is something like a guide grounded in philosophical thought.
These two experiences have changed the way I see the texts and ideas I have been teaching for years. How might students engage with them? What might they use them for? I now think more clearly about these questions. In other words, I think more about putting them to practical use. Vibe coding has taught me to think like an engineer.
There is an old joke: “To someone with a hammer, everything looks like a nail.” Some people take this as a criticism: not everything in the world is a nail! But I see a deeper truth in it. By taking up the hammer, we have become a nail-driving species. To use a hammer well, we also need to look at the world with an eye to where we might drive a nail.
Vibe coding works much the same way for me. Now, as I look at the world, all kinds of applications come to mind. A new line of thought is always running in the background: could I turn this insight into something useful in practice? Could I make that hypothesis into an application? I am not saying everyone should think this way. But AI offers this possibility to writers who want to experiment.
Most teachers in the humanities and the arts share a concern: are the reading, writing, and thinking skills of students who regularly use AI declining? There is also evidence that relying too heavily on AI for particular tasks can prevent students from acquiring important cognitive skills.
Schools and colleges should therefore retain some teaching methods that predate AI. Even I, an enthusiastic user of AI, have benefited from those older educational methods for many years. Homework and exams encourage students to make an effort. It is therefore reasonable to design some assessments so that students cannot use AI for every task. Having them write in class or take oral exams can help.
Students need to be steered away from the temptation to use AI as an instant shortcut for any task. Only then will they develop skills such as taking notes, shaping ideas, conducting research, and revising. These include the ability to learn how to learn (metacognition).
At the same time, the students I taught in Korea made an entirely reasonable request. Colleges should also teach students how to use the ever-changing range of AI tools. Because these tools are developing so rapidly, simply teaching students how to use a particular application is not enough. We must also develop their ability to think about which tasks suit AI and how to use it. We might call this a “meta-AI” skill.
My experience of using AI over the past two and a half years has suggested a way forward. We need to learn to converse effectively with AI. We also need to think like engineers. We need to ask: how can we test an idea using AI or other tools? How can we put it to use?
One place where I have tried this approach is an online writing course I have been developing over the past three years. There, we teach writing step by step. How do you choose a productive research question? How do you find the information you need? How do you construct an argument? How do you gather evidence to support it? How do you revise your conclusion when you encounter counterarguments and contrary evidence? How do you organise an essay? How do you seek feedback from others and revise again? All of these are part of the course.
We wanted to introduce students to writing as an exercise in thinking. This includes avoiding common errors in reasoning. We believed students needed to work through this long, sometimes tedious process themselves. That is how sophisticated thinking skills develop. Here, writing is the means of thinking. The work must therefore not be handed over to AI.
After college, students may rely on AI for many kinds of writing tasks. Even so, they need to develop the thinking skills that can be acquired only by writing a long essay themselves and struggling through repeated revisions. Only by learning this way will they be able to judge the quality of AI-generated text.
This experience prompted us to emphasise something else. Stylistic skills matter: boldly taking a new direction in writing, putting together unexpected phrases, creating unconventional metaphors. It is well known that AI is not yet particularly good at these things. AI-generated text may be adequate for many routine writing tasks. But there is a distinct value in knowing how to write better than AI.
At the same time, we realised that, used properly, AI can help us think and write. To that end, we teach students to use AI as a partner in debate. It offers counterarguments and contrasting evidence. Students put their ideas to the test. In this way, their thinking becomes sharper.
AI can also serve as a research partner. It can help us try out different ways of structuring an essay. To make our online course accessible to large numbers of students, we also created AI assistants that provide precise feedback on first drafts. They help students plan their revisions, too.
Since developing this course, I have been using AI in much the same way. I have created a few custom assistants that are familiar with my projects. I have asked them to rigorously test my ideas and arguments. This is certainly no substitute for reading what people have written or talking with them. But it has become an important critical practice in the way I think and write.
AI is a very powerful technology. It will have consequences both expected and unexpected. There are good reasons to be concerned about its impact on employment, education, and many other areas. That is precisely why those of us who work with words, too, need to understand AI properly.
What is it? How should we interact with it? We have literary theory and cultural history to help us seek answers to these questions. Let us use them.
