A reader put ChatGPT to the test with 90 questions on 88 authors

A reader has documented nine rounds of questions to ChatGPT about 88 literature authors, with ten questions per round and two authors repeated. The resulting log distinguishes between works read, started, in pogre, pending and merely known, and makes clear that the model improvised its answers without previously consulting any library.

There is a scene that repeats itself in many conversations about artificial intelligence and that almost never gets examined calmly: someone asks a language model about a book, the model responds with ease, and nobody checks whether that ease corresponds to anything. A reader has taken that scene into the territory of literature with a method worth a close look. Over nine rounds, he posed **ChatGPT** a total of **90 questions** about **88 different authors**. Two authors appeared twice, hence the gap between questions and names. The answers were given without previously consulting any library or reviewing bibliographies. The result is a log that separates what the model claims to have read from what it only knows by hearsay, what it started and abandoned, what it has pending and what it does not remember precisely enough.

The exercise is not an academic test, nor does it pretend to be. It is a personal inventory, ordered alphabetically, in which each author receives a label: read, started, reading, pending, known or unconfirmed. That taxonomy is the most interesting part of the matter, because it forces a distinction between having finished a book, having abandoned it halfway, having it on the shelf and knowing it exists. Most conversations about what a model "knows" of literature stop at the last rung and confuse it with the first.

## What it means for a model to "have read" an author

The first question this log raises is what on earth it means for ChatGPT to have read something. A language model does not read. It is trained on texts, and from that training it extracts language patterns that allow it to generate plausible answers about authors, titles and plots. When someone asks it about **Italo Calvino**, the model can list "Invisible Cities", "If on a winter's night a traveler" and the "Our Ancestors" trilogy without having turned a single page. What it does is reassemble information it has seen in reviews, articles, wikis and forums. That is not reading. It is something else, and confusing the two has consequences.

The log at hand deserves credit for not confusing them. It distinguishes between finished works and started works, between current readings and pending ones. When the model says it has read "Blood Meridian", "Suttree" and the border trilogy of **Cormac McCarthy**, it is asserting something a human reader could verify by browsing their bookshelf. When it says it started "The Brothers Karamazov" by **Dostoyevsky** and did not finish it, it is making a confession no model makes on its own initiative. The distinction matters because the everyday use of these systems tends to treat any fluent answer as consolidated knowledge.

There is a detail worth underlining: the log itself warns that the answers were improvised, without previously consulting the library or reviewing bibliographies. That means the model answered from memory, so to speak, and that some titles appeared with hesitations, corrections and approximate references. The reader had to reconstruct afterwards what the model had meant when it mentioned "a story related to the nights" or "a work related to music". That reconstruction is the human part of the experiment, and also the most revealing.

## The authors the model says it has read and those it only knows

The log reveals a pattern. There are authors about whom ChatGPT is expansive and confident: **Gene Wolfe**, with the "Book of the New Sun" tetralogy read three times according to the model, plus "Peace", "Castleview" and several stories. **Stanisław Lem**, with a dozen titles listed, from "Solaris" to "Fiasco", including "Cyberiad" and "A Perfect Vacuum". **Mervyn Peake**, with the "Gormenghast" trilogy read two or three times. **Flann O'Brien**, with four novels. **Lord Dunsany**, with five volumes. **Marcel Schwob**, with four books. **Stefan Zweig**, with eight titles and the warning that there are more stories it does not detail.

At the other extreme are the authors the model recognizes but does not claim to have read. **William Faulkner**, whose titles it knows, such as "The Sound and the Fury" and "Absalom, Absalom!", but does not count as read. **Kenzaburō Ōe**, whose works it has in digital library form but without confirmed readings. **Sándor Márai**, **Yasunari Kawabata**, **Joseph Roth**, **Robert Musil**, **Juan Carlos Onetti**, **Juan Benet**, **Kōbō Abe**, **László Krasznahorkai**. In several cases the model mentions the author's importance within their literary tradition, which sounds like encyclopedic knowledge rather than reading.

There is a third group, the most uncomfortable: authors the model did not recognize at all. **Anna Kavan** was not on its original list. Neither was **Vernon Lee**. **Jean Ray** sounded familiar as a writer of fantastic literature, but without confirmed readings. **Leopoldo Marechal** sounded vaguely familiar. **Ryūnosuke Akutagawa** was not clearly recognized. These cases are interesting because they show the limits of training: there are authors who are canonical in certain circles who simply do not appear with enough weight in the texts the model was trained on.

## The problem of translations and editions

One of the most curious aspects of the log is the attention it pays to editions. The model mentions that it remembers "The Mirror of the Sea" by **Joseph Conrad** for its translation by **Javier Marías**. That it has "Dhalgren" by **Samuel R. Delany** in an Ultramar edition whose cover was illustrated by **Antoni Garcés**. That it started "Portnoy's Complaint" by **Philip Roth** in an old Bruguera edition whose translation it found unsatisfactory. That it is concerned about the quality of some Spanish translations of **William Faulkner** and has therefore not seriously begun his work.

This attention to editions is more human than machine-like. A language model has no bookshelves, remembers no covers and complains about no translations. What it does is generate sentences that sound like someone has bookshelves, remembers covers and complains about translations. The log collects those sentences and orders them, and in doing so reveals how far the model is reproducing a plausible discourse about reading rather than a real reading experience.

There is also room for uncertainty. The model does not remember precisely whether it read "The Bloody Chamber" by **Angela Carter** or only started it. It cannot identify a **Gene Wolfe** story related to the nights, possibly "Seven American Nights". It does not know whether the English book by **Michael Moorcock** it has is "Mother London" or "London Bone". It does not confirm whether it read "The Golden Cockerel" by **Juan Rulfo**. These gaps are normal in any reader, and the log records them without dramatizing them. But they are also proof that the model is not consulting a verifiable internal database, but generating answers from patterns.

## What this exam says about using AI as a literary source

The practical conclusion of all this is not that ChatGPT lies. It is that ChatGPT generates plausible answers, and plausibility is not the same as truth. When a model says it has read **Clarice Lispector** but cannot specify titles, or when it says **Felisleto Hernández** bored it, it is producing statements that sound like reading experience but cannot be verified against anything. The log at hand deserves credit for having created a framework where those claims can be ordered and contrasted, at least internally.

For anyone using these systems as support in literary work, the lesson is clear: a model can help recall a title, place an author in a tradition or generate a list of possible readings. What it cannot do is replace reading. And when it claims to have read, it should be treated as what it is: a plausible hypothesis about what a human reader might have read, not a verifiable testimony. The log of the nine rounds, with its **88 authors** and its **90 questions**, is a good sample of how far that plausibility goes and where the terrain begins that can only be confirmed by opening the book.