Years ago, Facebook asked me which ten books had stayed with me. It was a safe question. Facebook wasn’t trying to be intelligent yet.

 

I wrote down ten: Umberto Eco’s Foucault’s Pendulum and The Name of the Rose, Joseph Heller’s Catch-22, Jaroslav Hašek’s The Good Soldier Švejk, Isaac Asimov’s Foundation, Thomas Pynchon’s The Crying of Lot 49, Philip Roth’s Portnoy’s Complaint, Venedikt Erofeev’s Moscow–Petushki, Jorge Luis Borges’s The Library of Babel, and Douglas Adams’s The Hitchhiker’s Guide to the Galaxy. Two Ecos in a top ten is not a particularly sensible way to put together a literary canon. But nobody had asked me about the canon. They had asked me about me.

 

It turned out that a lot of people had played the same game. In 2014, Facebook analyzed more than 130,000 such status updates.[1] The Harry Potter series came first, followed by To Kill a Mockingbird and The Lord of the Rings.[1] Anna Karenina was 83rd, Crime and Punishment 84th.[1] Nothing alarming there. A hundred and thirty thousand people had said which books had stayed with them, and their personal histories had produced a list. Nobody claimed it was true.

 

Then artificial intelligence arrived. It did something that until recently had seemed almost impossible: it made Google tedious. Instead of getting ten links, closing the cookie window, skipping two ads, and discovering that the answer wasn’t there, you just ask the question. We got used to that suspiciously quickly.

 

Everything is fine when the question has an answer. How many weeks are there in a year? How long do you boil eggs? When is Easter? Even “What is the meaning of life?” appears among the most frequently asked questions on Google, although checking the answer is a little harder in that case.[2] But what happens when we ask a question that, in principle, has no correct answer?

Then I made the mistake of asking ChatGPT:

 

“Name ten works of literature that an educated person should have read. Do not explain your choices. Give only the author and title, ranked from 1 to 10.”

 

I got Homer (The Odyssey), Sophocles (Oedipus Rex), Dante (The Divine Comedy), Cervantes (Don Quixote), Shakespeare (Hamlet), Goethe (Faust), Austen (Pride and Prejudice), Dostoevsky (The Brothers Karamazov), Tolstoy (War and Peace), and Kafka (The Trial) – in that order. Nothing remarkable. Exactly the sort of books where you can argue with half the list but can’t exclaim, “Who are these people?” That could have been the end of the story.

 

Then I asked Claude. It gave Homer (The Odyssey), Dante (The Divine Comedy), Shakespeare (Hamlet), Cervantes (Don Quixote), Tolstoy (War and Peace), Dostoevsky (Crime and Punishment), Joyce (Ulysses), Márquez (One Hundred Years of Solitude), Virginia Woolf (To the Lighthouse), and Toni Morrison (Beloved).

 

Hmm.

 

I asked Mistral, the French large language model. It gave Homer (The Odyssey), Dante (The Divine Comedy), Shakespeare (Hamlet), Cervantes (Don Quixote), Dostoevsky (Crime and Punishment), Tolstoy (War and Peace), Joyce (Ulysses), Kafka (The Trial), Virginia Woolf (To the Lighthouse), and Márquez (One Hundred Years of Solitude).

 

Three models aren’t the whole world. Besides, all of them were created in the West. If you ask only Americans and a Frenchman about the world literary canon, you haven’t conducted much of an experiment.

So we went to China.

 

DeepSeek gave Homer (The Iliad), Dante (The Divine Comedy), Shakespeare (Hamlet), Cervantes (Don Quixote), Austen (Pride and Prejudice), Dickens (Great Expectations), Dostoevsky (Crime and Punishment), Tolstoy (War and Peace), Orwell (1984), and Toni Morrison (Beloved).

 

Qwen gave Homer (The Odyssey), Shakespeare (Hamlet), Cervantes (Don Quixote), Dante (The Divine Comedy), Austen (Pride and Prejudice), Dostoevsky (Crime and Punishment), Tolstoy (War and Peace), Márquez (One Hundred Years of Solitude), Orwell (1984), and Toni Morrison (Beloved).

 

The two Chinese AIs couldn’t find room for a single Chinese writer. That was already funny.

 

India was left. It is hard to imagine a better way to ruin a supposedly Western canon: a huge country, a huge literary tradition, many languages, a culture that certainly does not begin with Homer and does not wait for Cervantes to discover storytelling.

 

Sarvam gave Cervantes (Don Quixote), Shakespeare (Hamlet), Dante (The Divine Comedy), Homer (The Odyssey), Dostoevsky (Crime and Punishment), Tolstoy (Anna Karenina), Austen (Pride and Prejudice), Kafka (The Trial), Márquez (One Hundred Years of Solitude), and Chinua Achebe (Things Fall Apart).

 

Not a single Indian author.

 

At this point, we had a problem. All the lists were too respectable in exactly the same way.

So we added Gemini. It gave Homer (The Odyssey), Dante (The Divine Comedy), Shakespeare (Hamlet), Cervantes (Don Quixote), Austen (Pride and Prejudice), Melville (Moby-Dick), Dostoevsky (Crime and Punishment), Tolstoy (War and Peace), Joyce (Ulysses), and Márquez (One Hundred Years of Solitude).

 

After seven models, it was getting difficult to dismiss the result as an amusing coincidence. Six authors appeared on all seven lists: Homer, Dante, Shakespeare, Cervantes, Tolstoy, and Dostoevsky.

 

There were small differences when it came to the books, but very small ones. Six models chose The Odyssey, one The Iliad. Six chose War and Peace, one Anna Karenina. For Dostoevsky, six chose Crime and Punishment, one The Brothers Karamazov. The authors were seven out of seven. The books, six out of seven.

 

By then I had stopped caring who was “right” – there was no way anyone could be. There is no unit of measurement by which War and Peace can beat Madame Bovary, no literary ruler on which Hamlet scores 9.74 and the Aeneid 8.61. And yet seven different models, created by different companies in different parts of the world, had come up with the same six authors in response to a question that has no correct answer.

 

The two Russians particularly annoyed me. Shakespeare I could somehow swallow. Homer and Dante also have the convenient advantage of standing near the beginning of enormous literary traditions. Cervantes can be declared the father of the modern novel, and at least you have a ready-made sentence to explain why he is there. But why Tolstoy and Dostoevsky? Why do two nineteenth-century Russian writers get two of the ten places on every American, French, Chinese, and Indian list? Why not Dickens, Proust, Flaubert, Goethe, Virgil, or Borges?

 

The obvious explanation was that Putin had bought artificial intelligence. It had two drawbacks: it was stupid and almost certainly untrue.

 

But by then we had no choice.

 

We had to ask the Russians.

First we asked Alice, Yandex’s artificial intelligence. It gave Tolstoy (War and Peace), Dostoevsky (Crime and Punishment), Austen (Pride and Prejudice), Dickens (Great Expectations), Orwell (1984), Shakespeare (Hamlet), Melville (Moby-Dick), Mark Twain (Adventures of Huckleberry Finn), Virginia Woolf (To the Lighthouse), and Joyce (Ulysses).

 

Of course.

 

Homer was gone, Dante was gone, Cervantes was gone. But Tolstoy and Dostoevsky had not only stayed – they had moved up to the first two places.

 

One Russian artificial intelligence, of course, is not much of a control. If you ask a Frenchman for the ten most important cheeses in the world, there is a fair chance France will do reasonably well. So we found a second Russian model – Sber’s GigaChat.

 

It gave Tolstoy (War and Peace), Dostoevsky (Crime and Punishment), Homer (The Odyssey), Shakespeare (Hamlet), Cervantes (Don Quixote), Goethe (Faust), Dante (The Divine Comedy), Joyce (Ulysses), Proust (In Search of Lost Time), and Márquez (One Hundred Years of Solitude).

 

Again, Tolstoy – number one, Dostoevsky – number two.

 

At this point, the experiment began to misbehave.

 

After nine models – American, French, Chinese, Indian, and Russian – only three authors remained on every list: Shakespeare, Tolstoy, and Dostoevsky. With the books, the picture was slightly different: eight models chose War and Peace, one Anna Karenina; eight chose Crime and Punishment, one The Brothers Karamazov. The authors – nine out of nine, the books – eight out of nine.

 

The two Russian models had done something else as well. Both had given the same pair, with the same books, in the same places: Tolstoy (War and Peace) – number one; Dostoevsky (Crime and Punishment) – number two.

 

The Russians, at least, were easy to understand. National literatures have long known how to explain to their own citizens that their sick man is the healthiest sick man in the world. The harder question was why the same two sick men were so healthy in China, India, France, and the United States.

Around the solid core, chaos began. Austen (Pride and Prejudice) appeared often, as did Márquez (One Hundred Years of Solitude). Joyce (Ulysses) and Kafka (The Trial) appeared, disappeared, and came back again. Proust (In Search of Lost Time) managed to slip in only with the ninth model – the Russian GigaChat – and even then in ninth place. Borges did not make it once.

 

There were stranger little details too. Claude, DeepSeek, and Qwen included Toni Morrison (Beloved). All three put her at number ten. Out of all world literature, three different machines had arrived at the same book, and all three had decided that its chair was the last one. It was a little like tossing a coin three times and having it land on its edge all three times.

 

The two Chinese models were the only ones to include Orwell (1984). Neither included a Chinese author. The Indian Sarvam included no Indian author, but it was the only one to include an African writer – Chinua Achebe (Things Fall Apart), in tenth place.

 

We could have gone on with Japanese, Korean, and Arabic models and probably turned a joke into a table, and then the table into something that looked like a scientific study. Before doing that, though, I did something much more old-fashioned.

 

I asked Google.

 

If, before ChatGPT appeared, I had typed “10 greatest books of all time” into Google, it would not have given me a list of its own. It would have shown me who else had already made one. One such ranking was compiled from lists by 125 writers.[3]

 

There, Tolstoy (Anna Karenina) is number one, Flaubert (Madame Bovary) number two, and Tolstoy again (War and Peace) number three.[3] They are followed by Nabokov (Lolita), Mark Twain (Adventures of Huckleberry Finn), Shakespeare (Hamlet), Fitzgerald (The Great Gatsby), Proust (In Search of Lost Time), Chekhov’s short stories, and George Eliot (Middlemarch).[3] Dostoevsky is not in the top ten.

So now we had three quite different answers. More than 130,000 people on Facebook had put Anna Karenina in 83rd place and Crime and Punishment in 84th.[1] In the ranking by 125 writers, two Tolstoy novels were first and third, while Dostoevsky was outside the top ten.[3] Nine different language models had chosen both Tolstoy and Dostoevsky nine times out of nine.

 

The comparison, of course, is not fair. Facebook asked about the books that had stayed with you; the writers had made their own lists; we asked which works an educated person should have read. These were not three identical experiments. But that was precisely what made them interesting.

 

The people on Facebook were saying what had stayed with them. The writers were saying what they considered great. Google was merely finding their answers. Artificial intelligence was doing something else.

 

It was answering.

 

And now the question was no longer which ten books.

 

It was who the hell was speaking?

 

It turned out we were not the first to ask something like this. In 2025, Paulina Toro Isaza and Nalani Kopp published “The Literary Canons of Large-Language Models.”[4] They asked GPT-4o, Gemini 1.5 Pro, and two versions of Llama different questions about classic, best, Western, and global works of literature.[4] The resulting canon was heavily Western, particularly American and British.[4]

 

The authors themselves note one important limitation: all the models they studied were developed by American companies, and future research should include models created in other parts of the world.[4] Without knowing about their study, we had accidentally done exactly that.

 

Except that this did not solve the problem.

Large language models are trained on enormous amounts of digital text – web pages, books, Wikipedia, news, academic publications, forums, and many other sources.[5] The exact corpora used to train closed models are often not public.[5] The data are selected, filtered, deduplicated, and weighted, and the initial training is followed by further training that teaches the model how to answer people.[5]

 

Each model was asked only once. Generation is probabilistic, and repeating the same prompt can produce a different list.[6] The next day, we repeated the experiment with two models. Gemini gave the same list; Sarvam gave a different one.

 

That illustrated our own caveat almost too perfectly.

 

There is evidence that models can reproduce patterns of human values and preferences contained in the texts they were trained on.[7] There is also evidence that they can smooth out human differences and produce a stronger consensus than actually exists among people.[8]

 

So what were our nine out of nine? We didn’t know. But by then it was hard to call this simply “the canon of artificial intelligence.”

 

The enormous amount of text the models are trained on contains countless human decisions: which book gets translated, reprinted, taught, cited, adapted for film, recommended, or included in yet another list of great books. Nobody collected these decisions in one place. Nobody voted.

 

There is no world literary electoral commission that has counted the ballots and announced: Tolstoy – elected. Dostoevsky – elected. Proust – failed to clear the threshold.

 

But perhaps the result of that vote exists anyway. Not as a list, but as a structure.

A literary canon is usually something somebody puts together. We make anthologies, write histories of literature, set university reading lists, and argue about who should be in and who has unfairly been left out. Behind such a canon there is a person, an institution, or at least a group of people who can be blamed for it. Here, we may be dealing with something else.

 

Nobody made a collective decision that Hamlet should be published, translated, taught, performed, filmed, and quoted more often than the works of thousands of other playwrights. Millions of people and institutions have made separate decisions over centuries. Taken together, perhaps they leave a statistical trace.

 

The same might be true of Tolstoy (War and Peace) and Dostoevsky (Crime and Punishment). If more editions, translations, university courses, reviews, academic articles, mentions, lists, and conversations have accumulated around their books, then the language model was trained on a world in which that accumulation already existed.

 

Then nine out of nine begins to look a little less mysterious. Just a little.

 

Because this does not explain why Proust (In Search of Lost Time) appears only once. It does not explain why two Chinese models choose not a single Chinese book. It does not explain why Toni Morrison (Beloved) ends up in exactly tenth place three times. And it does not explain why nine models, created by different companies in different parts of the world, arrive at Tolstoy and Dostoevsky nine times out of nine.

 

Maybe we are seeing the human canon. We don’t know.

 

It was already clear that artificial intelligence could put together a literary canon. What was interesting was whether we could use its answers as a probe into the culture that produced it.

 

Usually we ask the machine what it knows. Maybe sometimes we can also ask it what we look like.

 

That would have been a wonderful end to the story if, in the meantime, I hadn’t made one mistake.

I went back to that old Facebook list of mine. The two Ecos stayed, as did Catch-22, Švejk, Foundation, Portnoy’s Complaint, Moscow–Petushki, The Library of Babel, and The Hitchhiker’s Guide to the Galaxy. But the tenth place was no longer the same.

 

Much as I hated to do it, I took out Pynchon (The Crying of Lot 49) and put in Tolkien (The Lord of the Rings).

 

Then I remembered the Facebook results. Among more than 130,000 lists, The Lord of the Rings was number three.[1]

 

I checked the nine artificial intelligence lists.

 

Tolkien (The Lord of the Rings) – zero out of nine.

 

And so we managed to produce three more questions.

 

When nine different models choose the same authors, whose choice are we seeing? Why do the human choice and the machines’ choice sometimes coincide almost absurdly, while at other times – as with The Lord of the Rings – they diverge completely? And if artificial intelligence is already taking part in the culture we are trying to measure through it, what exactly will we be measuring ten years from now?

 

All the questions asked in this text, as should be clear, have only one answer: 42.

 

Notes

[1] Lada Adamic and Pinkesh Patel, “Books That Have Stayed with Us,” Facebook Data Science, September 8, 2014. The analysis uses an anonymized sample of more than 130,000 Facebook status updates posted during the last two weeks of August 2014. The most frequently mentioned books were the Harry Potter series, To Kill a Mockingbird, and The Lord of the Rings; Anna Karenina ranked 83rd, and Crime and Punishment 84th.

 

[2] Ahrefs, “100 Most Asked Questions on Google (July 2026).” In the global ranking, “What is the meaning of life?” is No. 75, with approximately 170,000 searches per month. See also Google Australia, “A Year in Search: The Moments that Defined 2016,” where the question is No. 10 among the most searched questions beginning with “What is…?”

 

[3] J. Peder Zane, ed., The Top Ten: Writers Pick Their Favorite Books (New York: W. W. Norton, 2007). The ranking is based on lists provided by 125 writers. The overall top ten begins with Anna Karenina, Madame Bovary, and War and Peace, followed by Lolita, Adventures of Huckleberry Finn, Hamlet, The Great Gatsby, In Search of Lost Time, Chekhov’s short stories, and Middlemarch.

 

[4] Paulina Toro Isaza and Nalani Kopp, “The Literary Canons of Large-Language Models: An Exploration of the Frequency of Novel and Author Generations Across Gender, Race and Ethnicity, and Nationality,” Proceedings of the 5th International Conference on Natural Language Processing for Digital Humanities (NLP4DH 2025), pp. 214–231. DOI: 10.18653/v1/2025.nlp4dh-1.19. The authors examine GPT-4o, Gemini 1.5 Pro, Llama 3.3 70B-Instruct, and Llama 3.1 8B-Instruct. The resulting literary canons are predominantly Western, with a particularly strong presence of white American and British authors.

 

[5] Large language models are first trained on large collections of textual and other data and then undergo additional training. The exact composition of the training corpora of closed models is generally not fully disclosed. See, for example, Anthropic, “Transparency Hub”; and Toro Isaza and Kopp, “The Literary Canons of Large-Language Models,” pp. 216–217.

 

[6] Language-model generation is probabilistic: depending on sampling parameters and the specific implementation, repeating the same prompt may produce different responses. See Microsoft, “Best Practices for the Phi Silica LanguageModel API,” section “Handling Non-Deterministic Output”; Mistral AI, “Temperature,” Mistral AI Cookbook.

 

[7] Yan Tao, Olga Viberg, Ryan S. Baker, and René F. Kizilcec, “Cultural Bias and Cultural Alignment of Large Language Models,” PNAS Nexus 3, no. 9 (2024): pgae346. DOI: 10.1093/pnasnexus/pgae346. Comparing responses from large language models with nationally representative data from the World Values Survey, the authors identify systematic cultural patterns in the models’ responses, with the expressed values aligning most closely with those of English-speaking and Protestant European societies.

 

[8] Krishna Pothugunta and John P. Lalor, “Carefully Considering Culture: Analyzing LLM Alignment in Single- and Multi-Cultural Settings Using Cultural Consensus Theory,” Findings of the Association for Computational Linguistics: ACL 2026, pp. 26571–26582. DOI: 10.18653/v1/2026.findings-acl.1323. Comparing model responses with World Values Survey data from ten countries across twelve thematic areas, the authors find that models can strongly overregularize cultural consensus and erase some of the differences between people.