Ulysses: Reading books together with an AI

Ulysses by James Joyce is considered one of the most difficult works in world literature. It is packed with countless, often subtle and hard-to-grasp allusions and references to mythology, politics, religion, current events, and other literary works. Between the chapters, a wide variety of stylistic devices are used, including the stream-of-consciousness technique, which makes reading even more confusing. I wondered whether ChatGPT might be a good tool to better understand this book—whether a joint reading experience supported by ChatGPT could be a wonderful complement that turns reading this work into a light and relaxing journey into Joyce’s world. I tried it out, and here’s what I found: is ChatGPT a good reading companion?

High-quality new editions of classics often offer not only a modern translation but also an extensive appendix with numerous annotations. These serve to provide readers with additional information and make the text and its background more accessible. There are books where this is a pleasant addition, but not strictly necessary. And then there are books like Ulysses by James Joyce. The text is so inaccessible and dense that the usual annotations completely fail, leaving the reader in the dark—with just a few marginal notes, there’s little help to be found.

There is a commented edition of Ulysses—a massive tome that provides explanations and notes on every page surrounding the original text. However, this makes reading even more exhausting. Wouldn’t it be much better to ask ChatGPT instead? After all, you can explicitly ask questions, request explanations for individual passages, or have entire chapters summarized. And ChatGPT should be quite familiar with Ulysses, since scholars have been analyzing and dissecting the work for over a hundred years. It’s safe to assume that several essays—and perhaps even the entire text—are part of ChatGPT’s training data. That might not be the case with lesser-known books, but Ulysses seems like the perfect candidate for such an experiment.

So, I read the first chapter, which I found quite understandable. The second chapter was already tougher, but still somewhat manageable. The third chapter, however, is completely confusing. In it, Stephen takes a walk by the sea, and the reader experiences his erratic thoughts, memories, and the walk itself—fully immersed in Joyce’s stream of consciousness. It’s utterly bewildering, and this jumble is almost impossible to untangle in a single reading. Chapter seven is also extreme, where Bloom moves among journalists and newspapermen, with numerous characters engaging in conversation without introduction. The tone of the journalists—casual, disrespectful, often cynical—is captured very well, similar to other novels that critically depict the profession. But the content itself is difficult to grasp. Some chapters, such as the sixth, “Hades,” are easier to follow when Bloom is actually out and about, but even then, one gets the feeling that much slips by unnoticed. There are simply many sections that are completely incomprehensible and utterly confusing.

When you start asking ChatGPT about the content of individual chapters, you do get usable answers. However, these are extremely condensed. Looking at the Wikipedia article for the book, you’ll find a detailed list of each chapter’s main themes and ideas. ChatGPT’s answers are roughly on this level when you simply ask about a chapter. You can easily ask, “What happens in the first chapter of Ulysses?” and get a short summary. That’s fine for the very difficult third chapter too—but it doesn’t really get you much further.

So I asked further: “Which foreign-language passages appear in the third chapter? Translate them for me.” ChatGPT listed three sample quotes that I couldn’t find in the text. When I asked for all French quotations, it gave me two (which I also couldn’t find) and claimed there were no others. Here, ChatGPT seems to fail—and when I compared it with a list I found online, the human research was clearly more reliable and complete.

Next, I wanted to know what the characters in the seventh chapter (the one with the journalists and their confusing dialogue) are actually talking about. Once again, the answer was very superficial and abstract, appearing incomplete and not particularly illuminating. When I asked for examples, ChatGPT referred to a few individual words, but overall remained at a high level of abstraction. For someone who understands almost nothing of that chapter, this offers no help or orientation.

In this way, I tried to extract more information from ChatGPT—and it turned out to be surprisingly tedious. You have to pull everything out of the AI bit by bit, and it becomes clear that it is optimized for providing summarized answers. Even though ChatGPT may seem too verbose in other contexts, when dealing with a text like Ulysses, its answers are far too brief. Detailed questions are often met with general, abstract statements such as: “It is important to note that ‘Ulysses’ is a very complex work with many layers and levels of meaning, which is why it is considered one of the most significant works of modern literature by many readers and critics.” Attempting to group chapters thematically and have ChatGPT explain sections only worked moderately well. For instance, the supposed visit to the aunt in the third chapter wasn’t mentioned at all, and later sections were explained so abstractly that it was difficult to see how they related to the actual text (the dog and other key elements were not mentioned at all).

I admit that Ulysses is a tough nut to crack—for human readers and therefore for AI as well. But ChatGPT’s explanations did not help me understand the chapters better or grasp their details. I also tried asking about individual passages—for example, “What is the fourth paragraph in the third chapter, and what is it about?” But soon, incorrect answers appeared, with wrong text excerpts that couldn’t be found in the book (not even when considering translation differences). I was never sure whether ChatGPT’s statements were actually correct. Since it never quoted directly from the text, it was impossible to tell where the information came from, whether it was complete, or even accurate. This, of course, is a fundamental issue with ChatGPT—one that becomes particularly obvious when reading a work full of dense, ambiguous passages like this one.

Reading Ulysses is hard work, and anyone who wants to truly understand the book rather than just skim through it must study it thoroughly. You need to go through it section by section, sentence by sentence, let it sink in, and bring along the necessary background knowledge to grasp the subtle allusions and humor. As a source or aid, ChatGPT seemed of little help here. If you want to go deep, you must ask very precise questions, continuously refine the answers, and still can’t be sure whether what you get is accurate. For a superficial overview, ChatGPT seemed fine—but even here, I found the Wikipedia article or a quick Google search much clearer and more helpful.

I should add that I find the current hype around artificial intelligence very exciting—and ChatGPT is a brilliant tool. Many things can be done wonderfully with it, and I’ve already connected the OpenAI API to a private project with excellent results. As with anything, there are use cases where ChatGPT works great, and others where it’s currently less suitable. You can also see how ChatGPT keeps improving and its knowledge base expanding. Just a few days ago, it was reported that ChatGPT now has access to the internet and is no longer limited to information up to September 2021. And these improvements are noticeable. Shortly after ChatGPT’s public release, I asked, “What races exist in Tad Williams’s Osten Ard?”—and the answer was complete nonsense. Now the question is answered correctly. It’s only a matter of time before ChatGPT is trained on more recent books as well. For example, I copied one of Selma Lagerlöf’s love stories into ChatGPT (in several chunks, since the input field is limited) and then asked questions about the story—and that worked wonderfully.

Conclusion: Of course, I made it too easy for myself and my expectations were too high. My hope was that ChatGPT would take a chapter and explain it in detail—what it’s about, clarify difficult passages on its own, translate foreign-language quotes, group and explain related passages. You can perhaps get this kind of information out of ChatGPT somehow, but it’s tedious and comes with the uncertainty of possible inaccuracies. Reading Ulysses therefore remains a challenging endeavor, and anyone seeking a deep understanding of the text will not get it handed to them—they’ll have to engage deeply with each chapter, research, consult additional sources, and bring a great deal of patience.