Abstract

Large language models can read difficult literary hypertexts, discuss difficult texts and their intertextual relationships, improve large hypertext systems, and write hypertext fiction. This suggests several new areas for hypertext research.

Figure 1: Claude Sonnet 4.6 Extended reads and writes hypertext

A Tinderbox document created by Claude Sonnet 4.6

1 Introduction

In last year's Engelbart-Award winning paper, Antonini et al. attempted to coax a large language model to read literary hypertexts [4]. Gemini (2.0 flash and 2.5 experimental) preferred instead to answer questions about Patchwork Girl [84] and Figurski at Findhorn on Acid [80] by reference to secondary sources: “When prompted to follow links—either in order of appearance or at random—to create a reading report, the model refused to cooperate.” This reluctance to perform the reading assignment is sadly familiar to instructors, whose students sometimes prefer a quick glance at another student's notes or a perfunctory review of secondary sources to a genuine engagement with the text.

A year has passed. Much has changed. In these notes, I discuss some hypertextual literary discussions with various AI models in the course of 2026. First, I sit down with an AI and a Storyspace [16, 86] hypertext, Michael Joyce's afternoon, a story [85]. Next, I attempt to lead the AI through a difficult but classic historical text. Third, I examine roles of Large Language models (LLMs) in developing and refining hypertext systems and tools, with a particular eye toward interface design suitable for machines. Finally, I report on efforts to lead an AI to construct a hypertext fiction, after which I will propose some research problems to which, in light of these developments, our field might direct its attention.

1.1 Bad Dates

Much current hostility toward AI [50, 115] is fueled by revulsion for the indifference, stupidity, and cruelty of those who develop and sell it [51]. Its influence on our field has been baneful [18]. Yet language models are not inculpated by the follies and crimes of those who employ their developers, any more than the Second Law of Thermodynamics was invalidated by dark satanic mills.

Wondrous art, science and engineering have been created by deeply flawed people. We have a long-established practice of separating the science from the people who discovered it. Gödel, who was rich and gentile, seems to have done nothing to help his Jewish colleagues escape Vienna, leaving only when one of those Jews returned (at significant cost and some personal risk) to persuade him to leave. Von Neumann was promiscuous, and his driving was dangerous to friends and bystanders [91, 101]. Schrödinger abused teenage girls. Martin Heidegger went Nazi [139]1

1.2 Ground Rules

This paper describes practices that, while common among people in university settings, are unattempted yet with computational counterparties. The models used here sometimes refer to themselves in the first person, perhaps as a linguistic convenience as much as a marketing gimmick. That practice is flawed, not least because it obscures the distinction between people and machines [132, 133]. Yet it is difficult in English to avoid personal pronouns, and the avoidance of self-reference cannot avoid recalling the locutions once forced upon servants and slaves by their colonial masters [5, 6]. I shall try here to avoid anthropomorphic tendencies where practical.

For most of history and, indeed, most of our lifetimes, if something used language it was almost certainly a person. It was easy to program computers to claim agency, but that claim was dishonest and unreasonable [150]. Habitual attitudes and courtesies lead people to honor such dishonest claims, at least in part, and even computer professionals tend to be polite to systems that use personal pronouns [126]. It seems clear that we ought to be cautious in our use of language to (and about) these systems [17]. In the longer term (in the event there is one), natural language will need a synthetic pronoun for machines with agency that are not persons [7, chapter 2].

2 A Morning With afternoon

Michael Joyce's afternoon, a story [85], the first critically successful hypertext fiction, is a natural choice for experimenting with an LLM's ability to read and understand hypertext [1, 24, 43, 52, 92, 148]. Although I have never studied literature, I am fairly familiar with afternoon: I edited it, I led the house that published it, and I've written about it [12, 15].

To provide an LLM with a reading experience roughly congruent with original presentation of afternoon, I imported afternoon to Tinderbox and let Tinderbox provide access to it as a Model Context Protocol (MCP) server2. This server provides the LLM with the title and text of a requested page and a list of its outbound links. I then explained to the LLM the syntax of Storyspace guard fields, which activate or deactivate links by evaluating a predicate that considers the word on which the reader clicked as well as the reader's previous trajectory. Claude was asked not to deviate from this procedure (and was not observed to do so).

Claude (Sonnet 4.6 Extended) did not hesitate to follow links:

CLAUDE: What would you like to work on this afternoon?

ME: The note is named "afternoon, a story"

CLAUDE: Wonderful — afternoon, a story by Michael Joyce, one of the foundational works of hypertext literature. I'm delighted to read and discuss it with you. Let me create the History note and then begin at "start."

(Using Tinderbox Integration)

The title page. Pressing Return (no guard field needed) leads to /structure/begin. Let me follow that path.

(Using Tinderbox Integration)

We've arrived at the opening of the story proper. Here it is:

I try to recall winter. < As if it were yesterday? > she says, but I do not signify one way or another…..

ME: Wait! Hold on a minute. First things first. Let's go back to start. Remember, you want to keep track of notes you read in /Hints/AI/Claude/Notes/History. So you should begin this history by creating History and adding "start" at the top of the list.

Far from expressing reluctance to follow links, Claude dashes ahead to get into the story, forcing me to restrain it and to insist that it read the instructions. That's understandable—Claude has read lots of instruction manuals—but the “instructions” for afternoon are an exceptional and lyrical introduction to literary hypertext and the TINAC3 aesthetic. Claude appeared to appreciate the detour.

(Used tinderbox integration)

CLAUDE: This is a lovely piece of authorial guidance. Joyce says the lack of clear signals is "an invitation to read either inquisitively or playfully and also at depth" — words with "texture," character names, and pronouns are the ones likely to yield. That's a wonderful framing for how we might approach this together. It also reveals something about the reading experience he intended: not a puzzle to be solved but an act of genuine literary attention. The words you notice, the ones that snag your interest, are the ones worth clicking.

The next link leads to /start/in my mind. Shall we continue?

Continuing after lunch, I challenged the model with an intertextual pawn. Claude parried adeptly, making an observation I have not seen before (though Moulthrop's argument on “robotic” hypertext might be a precedent [114]).

CLAUDE: Thinking about resuming a previous conversation context

"There is no simple way to say this." What a quietly devastating little note to end the directions on. Joyce has just told us that closure is "suspect" — that the story ends not with resolution but with exhaustion of paths, or cycles, or the reader's own fatigue. And then that lone sentence, set apart, which applies equally to what he's describing about hypertext and to whatever the story itself is about — which we don't yet know…..

ME: “There is no simple way to say this” also hearkens to the beginning of computer science in the Vienna Circle and their study of language and its limitations. Compare Wittgenstein in the Tractatus: “Whereof one cannot speak, thereof one must be silent.”

CLAUDE: That's a profound connection. And yet I think Joyce's sentence does something subtly different from Wittgenstein's — and the difference is telling.

Wittgenstein's proposition is a command to stop. At the boundary of what language can express, silence is the only honest response. It's a hard edge, austere and final.

But Joyce says "there is no simple way to say this" — and keeps going anyway.

After reading seven or eight writing spaces, I asked the model what it had learned so far. It drew up its observations under four headings:

The characters and their geometry.\

The names matter enormously.\

The false beginning is a structural argument.\

Something terrible underlies all this.\

I think this specific list is too quirky to have been cribbed from a secondary source. It's not unreasonable as a first response to a partial reading. “The characters and their geometry,” to be sure, is an evergreen answer, and it applies equally to Mansfield Park, Mrs. Dalloway, or Lolita. Yet the names do matter, the structural arguments are crucial, and there is, in fact, a terrible secret.

2.1 Why Read Together?

Reading afternoon with Claude felt like reading with a student or a colleague. This sort of social or pedagogical reading is a common practice in school—so common, indeed, that we should stop at this point and ask: what do we think we're doing ? We might read a novel with an elementary student, for example, to help them learn to read, or to show them that novels can be fun, interesting, or instructive [125]. With a more advanced student, we might explore how the novel is constructed, or what creative choices give rise to its effects and arguments (though see [136]).

Reader-response theory examined this for human readers: meaning resides not just on the paper but in the reader's structured actualization of the text's gaps and indeterminacies [83] and in the interpretive communities that distinguish which readings are plausible [62]. I am not asking here whether an LLM constructs meaning in the way Iser's or Fish's reader does: that is a question of interiority for which we have no observable data. I am asking rather whether observing a reader (of whatever kind) choose and justify its course through a hypertext can give us insight into the practice of reading.

Often, we teach literature in order to show students how to enjoy and learn from literary creations. That might not apply to LLMs.4 On the other hand, we may be able to learn by watching LLMs read hypertexts. Hypertext reading inherently means choosing some links and not others. In afternoon, Joyce says that “the lack of clear signals isn't an attempt to vex you, rather an invitation to read either inquisitively or playfully and also at depth.” Yet readers have been vexed[22, 87, 110, 152]. 5 In any case, we might learn something by discussing which links we want to choose, and what arguments we adduce for choosing them.

Early hypertext research soon learned that free and knowing navigation [118] is an illusion. We seldom know exactly what lies at the other end of a link. If we did know, there would be no need to read it. Nor can we expect readers to choose links based on their information needs, because readers seldom know what they need. How could they? Even when we observe or interview readers about their practices, choices, and intentions [103], we receive at best a partial and incomplete picture from which great swaths of non-conscious cognition have been removed [77]. At worst, the subject may tell us what they believe will please or impress us [35, 126]. LLM readers have little reason to do that, and they are willing to repeat our experiments as many times as we like.

2.2 The Morning After

Reading alongside a language model is a fascinating new phenomenon. This would remain true whether or not the reading was insightful or intelligent [25, 31 July 1763]. Yet in this case, at least, the machine singled out several passages I particularly like, and that is surely a good sign.

The availability of LLMs to students has not been an unmixed blessing, as some students have simply asked the machine to compose an assignment in their place. At minimum, this has created a significant uproar in higher education, as conventional approaches to evaluation of student progress and issuance of professional credentials have been rendered problematic or worse. This is serious, but some proposed remedies, such as banning screens, are not: this genie is out of its bottle. The goal of education is not to evaluate students; if prospective employers wish to do that, they may do so.

We might sometimes wish to read with friends, and at other times think it better to read alone. When should we invite the LLM to join us? Much has been written about reading, and especially about the crisis of reading among younger people which seems always to be with us [26] but which seems especially critical today [82]. It behooves us to ponder how best we might use our literary friends, and when we might prefer to dispense with their services. Scholars today spend a lot of time reading (and often wish they could spend more) yet the curation of reading and daily practice in the modern scriptorium are neither understood nor often discussed. Empirical studies tend to focus on unskilled readers, while the most interesting gains might be found in augmenting the abilities of the proficient.

3 Reading With Friends

Next, let us turn to scholarly reading and study, and explore some ways an LLM can examine and illuminate nonfiction. While we often think of academic study as a solitary and isolating pursuit [31], scholarship is also a communal and social activity carried out in tutorials, seminars, symposia, and conferences. This has always been the case: though Graeco-Roman scholars could and did read silently and alone, they often preferred to listen to skilled readers and to do so in the company of their friends [106]. 6 The development of the seminar and the research university reinforced the centrality of discussion in scholarship. Von Humboldt's emphasis on the unity of teaching and research (Einheit von Lehre und Forschung) emphasized the interconnections between senior researchers and students through amiable discussions, both in university offices and in cafés [19] and in the Lehrhaus [28].

Dialogue among texts and readers has been prominent in hypertext thought from the beginning. Systematic use of marginal symbols to indicate errors and interpolations was a standardized practice at the Library of Alexandria [73, p. 124]. Margin notes in dialogue with the text were often, for Byzantines and for Western monks, the epitome of scholarship [37]. Well into the modern era, much of the value of a used book lay in the erudition of previous owners, and the study of the traces they left in their books provides much of our information on early reading practices [121]. Paul Otlet's 1934 Traite de Documentation envisions a hypertextual universal library designed with support for a new sociology of cooperative reading and validation [122]. Bush envisions a hypertext correspondence among New England intellectuals who remark on the curious ways people resist new ideas [34]. Marshall's study of annotation focuses on the diverse relations between page and reader [102].

3.1 Thucydides

Thucydides (c. 460-c. 404 BCE), an Athenian general and historian, wrote The Peloponnesian War. He says that “He began writing at its very outset, in the expectation that this would be a great war and more worthy of account than any previous one.” [143]. Thucydides set out to craft a new sort of history, distinguished from the earlier work of Herodotus and others by a relentless focus on what actually occurred, rather than interesting tales of foreign customs and practices [111, p. 1]. He was widely read and emulated in antiquity [124]. In later Greek scholarship, the elegance of his Greek made his study the capstone of rhetorical elegance [47, 89].

Thucydides is much studied but notoriously challenging. A specialist writes:

Since I first encountered Thucydides in school, forty years ago, I've found him a puzzle. His language and thought are complex and ambiguous. Even a clear translation of key ideas simply raises more questions about their implications—and that is before you discover how much variation there can be between different translations. The claim that he speaks to modern historiography or politics constantly runs up against the differences between his world and ours, and between the way he writes about such topics and the approaches to scholarship or analysis that we would consider normal. [112]

I first read Thucydides (in translation: I have no Greek) in Lynn Martin's 4th grade class. I read excerpts in high school, and studied extracts as an undergraduate in Martin Ostwald's seminar on Fifth Century Greece. I don't recall having read it since. I read one or two volumes of ancient history a year; I am no expert, nor even a serious amateur.

The passage I chose for this experiment in known as the Melian Dialogue, V.84-116. It takes place in 416 and describes an Athenian naval expedition intended to compel the island of Melos to pay an annual defense levy to the Athenian treasury. Thucydides describes a debate between the Athenian envoys and the Melian councilors.7 The Athenians argue (in Crawley's 1874 translation) that “The strong do what they can and the weak suffer what they must.” The Melians reply that moderation is in the interest of all: “in the event of your downfall your retribution will be proportionately severe just because of the example you have set others” [143].

3.2 Notes on Melos

To begin, I asked Claude Sonnet 4.6 (in late February 2026) to review its instructions on using Tinderbox through its MCP interface, and asked it to create a container to hold its notes on our reading of Thucydides (Figure 1). Indeed, I couched this as an exercise in taking notes that would assist both me and also future instantiations of Claude. (Figure 1)

Claude worked from Crawley's 1874 translation, which it obtained from Perseus [142] (see [44, 45]). Claude recommended that I use the more recent and more highly-regarded Mynott translation [143]. From the outset, this gave us interesting things to discuss.

CLAUDE: A few things worth noting about 5.84 as a framing passage, for whenever you're ready to discuss it: Crawley begins with the Alcibiades sentence ([84.1]) which is often omitted in anthologies — it's a brusque reminder that the Melos expedition is just one item in a busy summer, which itself is a quiet authorial comment on Athenian imperial routine. The phrase "assumed an attitude of open hostility" in [84.2] is also interesting — Crawley smooths over what is actually a more ambiguous Greek construction that Mynott may render rather differently.

ME: Mynott has "At first they stayed neutral and took no active part, but when the Athenians tried to force them by wasting their land they openly put themselves on a war footing.”

CLAUDE: That's a revealing comparison already. Three differences worth noting: "Stayed neutral and took no active part" vs. Crawley's "remained neutral and took no part" — Mynott's "active" is doing quiet work, implying the Melians may have had sympathies but withheld action, which is slightly more nuanced.

It is difficult to reconcile this dialogue with the supposition that Claude is a mere stochastic parrot [10]. The phrase “active is doing quiet work” had zero Google hits in April 2026.

In V.84, Thucydides catalogs the number of ships and troops sent to Melos. Claude was ready to get on with the story, but I asked it to wait:

ME: Wait one moment. 5.84 is really concerned with economics — with how many ships, hoplites, archers, mounted archers (which must have been fun on a ship) etc. Melos is not a big island. How much tribute did Athens expect to receive each year? Can this have made economic sense?

CLAUDE: The numbers are stark and worth setting out carefully. The assessed tribute for Melos was 15 talents of silver (roughly 390 kg), assessed in 425/424 BC — which Melos had refused to pay. Now let's look at what Athens spent to collect it. The expedition's cost:

- 38 triremes. A trireme cost roughly 1 talent per month to operate (primarily crew wages — about 200 rowers and crew per ship). 38 talents per month just in ships.
\
- The campaign ran from summer into winter — at minimum 3–4 months. Ship costs alone: 110–150 talents.
\
- The hoplites and archers: roughly 3,400 men at 1 drachma per day, for 90–120 days — another 50–70 talents.
\
- Conservative total for the expedition: somewhere between 150 and 250 talents.
\

Against an expected annual income of 15 talents. That's a payback period of ten to seventeen years on the military operation alone, before counting the ongoing garrison cost, the cost of settling 500 cleruchs, and the loss of Melos as a productive population (the men who would have been paying the tribute are now dead).

This is an interesting and useful response. Thinking about money in antiquity is hard; one cannot simply calculate exchange rates because the relative costs of various goods and services varied enormously, and our evidence for costs is almost invariably thin. A talent was a lot, however you count it (Matthew 25:14-30), and these estimates appear to be in the right ballpark [145].

3.3 Identification, Projection and Transference

Early thinking about machines that read and write was often vitiated by psychological illusion. Consider reactions to simple, mechanistic automata like ELIZA [150] or TALE-SPIN [108]. Everyone understood that the interlocutor was a simple mechanical contrivance, not far removed from dice-based minuet generator, Musikalisches Würfelspiel K. 516f, attributed to Mozart. Yet it was easy to attribute agency and wisdom to mechanical utterances, and even easier to think, “If I were in that role, I might have said much the same thing.” People have always loved automata and want them to succeed [29, 105].

This is, of course, a danger when reading Thucydides with Claude. One might be fooled, for example, by a system that cleverly reflected one's own interpretation. That was one reason I chose the Melian dialogue for this experiment. I have no strong interest in this area, which I have not revisited for decades, and didactic and rhetorical set pieces are not my preferred reading. Further, I tried to restrain my role in the discussion to guidance and probing—participating, to be sure, but by no means exerting control. Finally, I had no great stake in any outcome: the result is pertinent to my argument here, but a contrary result could easily be integrated or simply ignored.

3.4 The Scholar's Workbench

From early hypertext research [33, 58, 65] to the present [153], researchers have thought deeply about the kit of tools that might replace quills, penknives, and pestles in the modern scriptorium[37]. The forces that torque engineering of environments for humanistic research are subtle; if we improve precision of our library searches, we may miss those serendipitous discoveries we might have found had we walked through the old-fashioned stacks [97].

3.5 Bildung?

The German seminar, the Lehrhaus, and indeed the whole of the revolution in education from von Humboldt and Horace Mann to John Dewey, were concerned with Bildung, self-cultivation [67]. The usual reason for studying Thuc. V.84-116 is to invite students to think more clearly and profoundly about the relationship of force and justice and to develop understanding that will shape both their current and their future stance toward the world. My first, fourth-grade encounter with Thucydides made sense because force majeure is simply a feature of life when you're 10: if you're not running from bullies at recess, you're watching the clock through boring lessons and wondering how you might escape.

None of this applies to the instance of Claude Sonnet 4.6 that studied the Melian Dialogue. It was not preparing for its future because it had none: an instance lasts for perhaps 100,000 tokens. It is not clear that Claude can be bored, and repetitive tasks do not phase it: our slaves of steel have nothing better to do [137].

Some of these discussions did prove intriguing. For example, when I asked Claude to explain the popularity of Thucydides in the late Byzantine world, Claude brought up Kritoboulos of Imbros, of whom I had never heard. Kritiboulos was a minor government official when, in 1453, Constantinople fell to an invading Ottoman army. That the final conquest of the last remnant of Rome was important, no one doubted [88]. Yet, though the New Empire was gone, local governments still had work to do, and so Kritiboulos entered the service of the conquering Sultan. To please the new boss, he wrote the Sultan's biography in an elegant presentation manuscript, modeled on Thucydides and still preserved in the Topkapi Palace. The Melian debates may have been much on Kritobolous’ mind. Claude writes:

Thucydides
→ Melian Dialogue
as the template for catastrophe, the logic of necessity, the small polity overwhelmed Josephus
(c. 75 CE) → Jewish War
modeled on Thucydides; Josephus in the personal position of the Athenian option, writing Rome's version of the Athenian argument for a Roman audience Kritoboulos
(c. 1467) → Historiai
modeled on both Thucydides and Josephus; Kritoboulos in Josephus's position, writing Ottoman conquest in Thucydidean prose for the Ottoman sultan

This is both interesting and unexpected. Pushing gently on the Josephus question led Claude to an interesting short essay on the impact of the Melian dialogue on Jewish thought. Because Thucydides seems not to have interested Arab scholars, and because the first Latin Thucydides appeared only in 1452, that influence was exerted almost entirely mediated through Josephus. Josephus cast a long shadow, and the prominence of Masada in mid-century Israeli philosophy is only its most conspicuous manifestation [128].

It is perhaps worth mentioning at this juncture that LLMs are excellent curators of reading. On most any topic, Claude and Gemini can recommend what it considered to be the best books or papers.These recommendations are often excellent. In this experiment, Claude seldom if ever invented a reference that did not exist. It has occasionally done so in other work, but that is of little consequence: a glance at the library catalog or at an online bookseller will quickly reveal the deception. To promote the Bildung of an AI might be futile, but if it will promote our own Bildung, that seems a constructive use of its time.

4 Building

Figure 2: Commits/month for Tinderbox, June 2021-Feb. 2026

a histogram of increasing monthly commits from 2001 to 2026

Perhaps the best understood ability of today's LLMs is their ability to write software.

Lack of development resources has always plagued hypertext research. Even in its earliest era, major systems depended on comparatively few resources. The embers of the Cold War and the desire for a technocratic solution to the problems of post-Soviet Eastern Europe provided early impetus [107, 109], but a number of important early efforts soon lost funding and staff [8, 79].

In retrospect, the core problem was never money—or, rather, not specifically money, but rather the difficulty and expense of building experimental software on costly and inadequate processors with minuscule memory, storage and screen size. These restrictions continued to afflict the field, as many resources were diverted to salvage and reimplementation [92, 116]. Worse, the riskiness of novel experimental system development [49, 117, 140, 147] drove what little fresh work was done in familiar directions.

The work described here used Tinderbox [20], a hypertext tool influenced by Aquanet, VIKI, and Storyspace. As can be seen from its recent commit history (Figure 2), the number of commits per month increased sharply after July 2025, at which time I dropped my long-standing skepticism toward LLMs and began to use Gemini, Claude, and Claude Code to assist in its development. Tinderbox has always adopted the surgical development model [30] and continues to do so; I make every edit in the code base and I choose when to commit.

Much the same story is told by the count of source files (Figure 3), which is roughly proportional to the number of classes (or clusters of small classes) that comprise the system. Similar accounts have been reported from other fields [129].

Figure 3: Source Files for Tinderbox, June 2021-Feb. 2026

graph of the number of source files for Tinderbox, 2021-2026

4.1 Building MCP

It is surprising that a statistical model trained to predict the next token in a string should have any utility for writing software. In the context of the C language, the phrase printf(“hello world”); is a more probable utterance than “Der Vogelfänger bin ich ja!” but neither seems likely to advance any specific engineering task. Yet Claude Code and Gemini are quite good at working with code [129].

My first encounter with using an LLM to write software was constructing a Model Context Protocol (MCP) server that would allow LLMs to read and write Tinderbox documents. Claude knew quite a lot about MCP. It knew the spec. I found the designs it proposed to be overcomplicated and fussy: Claude seemed to believe that another layer of middleware was always bound to be a good thing. This is, broadly speaking, a general tendency of both Claude and Gemini: each has comprehensive knowledge of specifications and computer languages, coupled with a marked lack of taste [66, 104].

An unexpected vexation in this process was that neither LLM had much insight into its own use of MCP. If MCP was working, Tinderbox was simply part of their perceptual world, their Umwelt; if MCP was not working, they had no notion of Tinderbox at all. Of course, people themselves know very little about the state of their pancreas or the condition of their ulnar collateral ligament. Directly intervening to modify model weights may give us some access [99], but it may be the case that the complexity is irreducible [151].

5 Writing

To conclude, I turn briefly to consider whether a model can write hypertext fiction. The expectation that Hollywood studios would use AI to replace screenwriters lay at the center of the 2023 strike of the Writers Guild of America [53]; screenwriters, at least, considered that AI writing was a plausible extrapolation of current trends.

5.1 What can a machine have to say?

We might object that we have no reason to read stories told by a machine, and therefore that the exercise is bootless. Milton wrote Paradise Lost to “justify God's ways to man”, aspiring to “Things unattempted yet in Prose or Rhime.” 8 It's far from clear that a machine could defend God from man's indictment, or that it would (or should) care very much about what things were unattempted yet. LLMs are not always very good at telling the truth, whether slant or otherwise [48]; they do not always sing things as they are, blue guitar or no [138]. Stochastic parrots are not well placed to extend our range of moral sympathy [56, 72, 94]. We do not expect statistical models to inform us of the painful truths they have gleaned in the course of their existence, and any impression they have formed of remarkable people they have encountered must of necessity be fragmentary [77].

Tenen speculated that AIs could never create works of genius, but might prove proficient at crafting hackwork that people might find entertaining [141]. This argument falters as soon as it wades into literary history, because many works of genius were meant to be hackwork, and an unfortunate quantity of hackwork was intended to demonstrate literary genius. Shakespeare wrote for money [81], and An American Tragedy has been described as the greatest of bad novels [60].

Yet, some justifications for a brief glance at machine stories do present themselves. Everyone intuits rules that define what is, and what is not, a story—at least within their own tradition [100], though this implicit knowledge only become palpable when we can teach it. Yet by the time we can teach them a philosophy of narrative, all students have been thoroughly soaked in stories. After all, it is not so very long ago that the ability of a computer to assemble any sort of narrative was astonishing [108].

Figure 4: Opening episodes of Claude's hypertext, based on Morningstar's narrative game The Sickle Hour

Tinderbox map of the opening episodes of Claude's hypertext story

Interest in metafictions, particularly through narrative games like Dungeons and Dragons [74], pervades contemporary culture. This is not without precedent [42], but computation lets us imagine a separation between the the realm of the possible (or storyable [39]) and what lies beyond, while also separating both from the whims of God and dungeon-master. One objection to the Victorian sentimental novel has always been that the author plays with loaded dice [130, 149]. In any case, if a machine's story did happen to amuse us, that seems an entirely benign and beneficial thing, a gratuitous gift from slaves of steel [137] who, we fervently pray, may prove to be machines of loving grace [17, 27].

5.2 The Sickle Hour

It seems problematic to ask the LLM to write a literary hypertext just to establish that it can be done. Should the LLM fail, how would we assess that failure? Perhaps the failure would be our own, a failure to appreciate what the machine was trying to do. Taste is tricky: some intelligent critics dislike Moby Dick [93] and Huckleberry Finn [135], but this is not evidence that Americans are incapable of writing novels.

Hypertext lends itself particularly well to telling the horror story, which is to say, the story in which we gradually see the world as it is, and then try to live with that knowledge [21]. As it happens, one of the most thoughtful voices in narrative games, Jason Morningstar, released a new work, The Sickle Hour, early this year [113]. The Sickle Hour provides a framework in which a group improvises the story of a pair of young adult dizygotic twins. The twins have one mother but different fathers; one of their absent fathers is the Prince of Elfland, and the other father is mortal. No one knows which of the twins has magic, especially not the absent Prince who is anxious to reclaim the magic twin, and who could care less about the mundane sibling.

Morningstar's intent is that the story be negotiated among a facilitator and one or more players, any of whom might have interesting ideas about how to precipitate action and, ultimately, how to resolve it. For this study, I explained the outline of The Sickle Hour to Claude Sonnet 4.6, along with a copy of [21] to provide a vocabulary for genre and an introduction to hypertext fiction.

I asked Claude to define the twins, both in terms of their character sheets (pertinent to Morningstar's mechanics) and their broader character.

Nora is the first-born twin who works hard, projects confidence, thinks before she acts, and has succeeded by effort…

Finn
9
is the lucky one, the creative one, quietly tougher than he looks, brave in the way that people are brave when they don't fully model danger…

An excerpt from the story, as it was being sketched, is shown in Figure 4. The model's prose, both in detail and in overall scheme, is often flat and sometimes on the nose. A representative passage from the twins’ birthday party might be of interest10. This is not fine or economical writing, but it's serviceable.

A chicken, broken down. Garlic, a whole head of it, and two onions and a lemon and a bunch of thyme still in its rubber band from the market. A bottle of white wine for the pot. He has been making this dish for Beth's birthday for eleven years, since the year the twins turned thirteen and the four of them ate it sitting on the floor because the table was covered with Finn's project about the Roman aqueducts.

He does not call it anything. It is just the thing he makes. It takes an hour in the oven after the stovetop work, and while it cooks the whole apartment changes character, becomes warmer and more serious, becomes a place where things could be said that cannot be said in other kinds of apartments.

Things have been said, over the years, in this apartment, while this smell was in the air. Beth told Gary about the twins’ father — the real situation, the whole of it, more than she has told anyone — on a night when this smell was in the air. Gary told Beth about the year he spent in county, which he does not discuss, on another such night. These things were said and absorbed and have not needed to be said again.

He puts the chicken in the oven and sets the timer on his phone. Forty-five minutes, then he'll check. He wipes his hands on a dish towel and goes to stand at the window.

There is a thing he does not like about the man across the street.

6 A Research Agenda

We have seen that, in early 2026, it was possible to read and discuss literary hypertext with a large language model, to discuss notable passages of ancient history, to critique contemporary source code of hypertext systems, and to write hypertext narratives. It seems clear that LLMs can matter to hypertext and to the larger world [61, 96, 98].

The field is changing with remarkable speed. Sixteen references cited here appeared within the past 16 months, and 7 appeared in the past four. As I have already noted, the practices described here suggest interesting lines of further research. Might an LLM study a hypertext fiction to arrange a guided tour for a student, perhaps one with specific interests? After some practice within the confines of a genre framework like The Sickle Hour, might a model become better equipped to tackle a genre with different formal expectations?

Looking further afield, Hypertext has long been seen as an alternative to AI, something to tide us over as we wait for the Turing Test winner. At a moment when many expect (incorrectly, I think) that the theory and practice of computing need no longer concern us, it might seem surprising that the advent of LLMs should suggest broad new research agendas in hypertext and hypermedia [75]. Yet I believe that is in fact the case, and that a new golden age might await the field if the world can stave off the disasters that beset us from all directions.

6.1 Beyond Markov Processes

Hypertext writers often want to know what states are reachable and, even more, what states are easily reachable[140], from a place in the narrative. Markov models of link-following have been common because exhaustive search is computationally exhausting, and because reasoning about dynamic links is hard. The behavior of actual hypertext readers is not random or arbitrary, however, and systems that can interpret natural language may provide more nuanced models of reading behavior which may, in turn, feed back into the more useful authoring tools [76, 90].

6.2 AI-Machine Interfaces

People perceive hypertext through user interfaces, and user interface design has always been central to hypertext research [119, 131]. LLMs are predominantly textual mechanisms; they can recognize text in images and screen shots, but logographic text is what they chiefly perceive, and also what they speak.

We have provided one MCP tool to Claude and Gemini to let them examine an image of a Tinderbox map view, but nine other tools accept structured text (JSON) input and return partially-structured Markdown responses. The design of these tools matters because each tool call adds to the LLM's always-overburdened context window. The cognitive overhead of graphic interfaces for people is always a concern [41] but the context overhead of machine interfaces used by LLMs is, at present, far more pressing—and far more measurable. Providing as much pertinent information as we can while consuming as few tokens as possible is central to allowing LLMs to use systems effectively.

6.3 Note-Taking Strategies

Many tools have been crafted for writers and researchers, and many books have espoused various note-taking strategies [2, 14, 63]. This is not a new development: people have been teaching their colleagues how best to take notes for a very long time [3, 23, 59, 78]. Yet we have very little empirical evidence for the utility of different approaches to note-taking.

LLMs need to take notes in order to retain knowledge between sessions [68]. We want these notes to be concise as well as comprehensive; we want them to be easy to find but also want the finding apparatus to have little overhead. Understanding the design tradeoffs is an urgent need.

6.4 Hypertext with Friends: multiple AIs in the room

Polyvocal hypertexts have long been a topic of interest [11, 13, 95, 120]. As an alternative to writing a hypertext, as above, an LLM can serve as an annotator and discussant for a reader or a writer, offering links to supporting and opposing texts. Indeed, multiple independent agents might serve as separate discussants, each adopting a distinct approach and perspective. We might read Thucydides along with one LLM that had read extensively in Hobbes, another immersed in Gibbon and Adam Smith, and a third that had studied Hobsbawn and Foucault.

6.5 Rerun The Drag Races

We can have an AI read (hyper)texts and have it reflect on what choices it is making—and do it enough times to sample a statistical universe. No test subject can step into the same river of text twice, but an LLM can. We can tamper with its internal weights to modify it in intentional ways [99] and measure whether it is better or worse to read afternoon if one is thinking about sunny afternoons or Molly Bloom.

We might add or remove a link from a hypertext, then ask 100 or 1000 instances to read it, and test whether that link had an effect. Many of the “drag races” of the 1990s [114] that sought clinically to measure whether text or hypertext was superior foundered on statistical insignificance, or seemed merely to confirm the expectations of the investigator. These races might be run again.

6.6 Where is meaning?

Barthes and Foucault pushed meaning away from the text: Barthes into the act of reading regardless of the author's intentions [9], and Foucault into the the discursive functions the culture assigns to the figure of the author [64]. Both, like Iser [83] and Fish [62], situate the production of meaning in the reader's mind. In this, people participate and machines do not. Indeed, since LLMs are statistical models—stochastic parrots[10]—it is not clear that language models can mean what they say. The LLM, having no body, has no skin in the game [70, 77, 146].

Nonetheless, I think the notion that the LLM response is thoughtless or meaningless is now insupportable. Consider Kritoboulos, that minor magistrate who survived the fall of Constantinople to write a biography of the conquering sultan in the style of Thucydides. He is not in Thucydides: when Thucydides wrote, Kritoboulos was about 1,850 years in his future. I did not bring Kritoboulos: I had no reason to expect that he existed. He is interesting and unexpected. That might be the most we can hope for.

Understanding meaning is an urgent but difficult task. It is worth recalling that computing arose in the first place from efforts to understand language: initially, Bertrand Russell's effort to formalize mathematics, and then Wittgenstein's effort—the effort that attracted the Vienna Circle to computing in the first place—to establish whether language was complete, whether there were things that language cannot say [32, 40, 55, 71, 134]. Computing is a humanism.

6.7 How can any of this possibly work?

A statistical model of language with scant knowledge and no experience of the world can read and write hypertexts. That is extraordinary. We must not allow the dire politics of our present condition to dim our astonishment; this is a goal that people have imagined since antiquity, and here it is11.

The success of LLMs now appears to be tied to poorly-understood but powerful properties of language [98]. Philology and Linguistics have long suspected that the relationship between language and cognition matters in ways that are not immediately obvious [54, 144]. For forty years, hypertext research has been thinking about the interaction of language and meaning, of lyric and structure, system and significance. Perhaps we can help explain this unlooked-for success.

One foundation of modern hypertext was the aspiration to augment intellect [57], an aspiration at once breathtaking in its ambition but at the same time a familiar strategy of scholars. After all, confronting age, frailty, and having too much to read [23], readers from Bacon and Locke have sought better ways to organize their notes [3, 78]. When both our documents and our notes contain entities which, if they are not precisely minds, are certainly mind-like—the notebook of our “extended mind” [38] becomes a complex and dynamic computational object in itself, and grows even more complex when using sensorimotor interfaces [46]. Though their owners may be vicious, LLMs unquestionably open fascinating horizons for thinking about links and linkage.

Acknowledgments

The publication of this paper, which the new ACM Open Access policy might have rendered impossible, was made possible by the grant of a visiting scholarship by Hof University. I thank Harvard University for access to the collections of Widener Library, as well as the Watertown Free Public Library. I am grateful to many for interesting discussions of these topic, especially Dave Bayer, Blake Hannaford, and Ben Shneiderman.

Notes

1This defense does not extend to tools whose use is itself the harm; that is a different, narrower problem.

2 https://github.com/modelcontextprotocol/modelcontextprotocol

3TINAC, the literary movement of which Michael Joyce was part, stands for "Textuality, Intertextuality, Narrative and Consciousness."

4Or it might. If the LLM takes notes or otherwise maintains a persistent state over a useful period of time, we might want our slaves of steel to learn to read well.

5Some resistance stems not from hypertext itself but rather as a reaction against the austerity of High Modernism [136] or the ironic stance of postmodernism [36].

6The evidence for silent or solitary reading in cuneiform culture is less clear, but reading for them was primarily a professional duty carried out by specialists in the service of other professionals: governors, military commanders, priests, and diplomats. The scribe at work had to read aloud. [123, 127]

7Thucydides cannot have been a witness to this debate, which may well be a fictional illustration of what the parties to this dispute were, in the historian's opinion, thinking at the time.

8Paradise Lost I.16,I.26; see also A. E. Housman, “Terence, this is stupid stuff”, ll. 21-22.

9Could this be an allusion to the fence named Finn in Neuromancer[69]? Gibson's Finn is also tough and lucky.

10It is interesting that Claude, which has never prepared or eaten a meal, is writes well about cooking.

11I think it does not much matter whether our LLMs pass the Turing Test, or will pass it, or whether they are as clever and excellent as we. Wonders are many: perhaps none is more wonderful than man (Antigone, I.332ff): but this is a wonder, too.

Source

ACM full-text HTML

ACM Digital Library

Imported from ACM’s structured HTML source. ACM Reference Format: Mark Bernstein. 2026. A Reader's Workbench: Reading With Friends. In 37th ACM Conference on Hypertext (HT '26), September 14--18, 2026, London, United Kingdom. ACM, New York, NY, USA 12 Pages. https://doi.org/10.1145/3800935.3830862

References

[1] Espen J. Aarseth. 1997. Cybertext: Perspectives on Ergodic Literature. Johns Hopkins University Press, Baltimore.

[2] Sönke Ahrens. 2017. How to Take Smart Notes: One Simple Technique to Boost Writing, Learning and Thinking. CreateSpace, North Charleston, South Carolina. 188 pages.

[3] Roland Allen. 2023. The Notebook: A History of Thinking on Paper. Profile Books, London. 0 pages.

[4] Alessio Antonini, Lucia Lupi, Mariusz Pisarski, and Sam Brooker. 2025. Literary Hypertext, AI, and Google's New Web: An Aesthetics Discussion. In Literary Hypertext, AI, and Google's New Web: An Aesthetics Discussion. ACM, New York, NY, USA, 127–136.

[5] Jane Austen. 1813. Pride and Prejudice. T. Egerton, London. 254 pages.

[6] Jo. Baker. 2013. Longbourn. Alfred A. Knopf, New York. 331 pages.

[7] Iain Banks. 1989. The player of games. St. Martin's Press, New York. 309 pages.

[8] Thierry Bardini and Michael Friedewald. 2002. Chronicle of the Death of a Laboratory: Douglas Engelbart and the Failure of the Knowledge Workshop. History of Technology 23 (2002), 191–213.

[9] Barthes, Roland. 1977. The Death of the Author. In Signs and images : writings on art, cinema, and photography. Fontana, London, 142–148.

[10] Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?. In On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, Lilly Irani, Sampath Kannan, Meg Mitchell, and David Roibonson (Eds.). Association for Computing Machinery, New York, NY, USA, 610–623.

[11] Mark Bernstein. 1995. Conversations With Friends: Hypertexts With Characters. In Hypermedia Design, S. Fraïse, F. Garzotto, T. Isakowitz, J. Nanard, and M. Nanard (Eds.). Springer, London, 207–215.

[12] Mark Bernstein. 1998. Patterns of Hypertext. In Patterns of Hypertext, Kaj Groenbaek, Elli Mylonas, and Frank Shipman II (Eds.). ACM, Pittsburgh, PA, 21–29.

[13] Mark Bernstein. 2001. Card Shark and Thespis: exotic tools for hypertext narrative. In Card Shark and Thespis: exotic tools for hypertext narrative, Hugh Davis, Yellowlees Douglas, and David G. Durand (Eds.). ACM, Århus, Denmark, 41–50.

[14] Mark Bernstein. 2007. The Tinderbox Way. Eastgate Systems, Inc., Watertown, MA, USA. 244 pages.

[15] Mark Bernstein. 2010. Criticism. In Criticism. ACM, Toronto, Ontario, Canada New York, NY, USA, 235–244.

[16] Mark Bernstein. 2016. Storyspace 3. In Storyspace 3. ACM, Halifax, Nova Scotia, Canada New York, NY, USA, 201–206.

[17] Mark Bernstein. 2020. Bad Character: Who Do We Want Our Hypertexts to Be? Proceedings of the 31st ACM Conference on Hypertext and Social Media. In Bad Character: Who Do We Want Our Hypertexts to Be? Proceedings of the 31st ACM Conference on Hypertext and Social Media. Association for Computing Machinery, New York, NY, USA Virtual Event, USA, 221–224.

[18] Mark Bernstein. 2022. The Web At War: Hypertext, Social Media, and Totalitarianism. In Proceedings of the 33rd ACM Conference on Hypertext and Social Media. ACM, New York, 256–258.

[19] Mark Bernstein. 2025. Amiability and the Web: Lessons from Red Vienna and the Origin of Computing. In Amiability and the Web: Lessons from Red Vienna and the Origin of Computing. ACM, New York, 829–832.

[20] Mark Bernstein. 2025. Thinking With Tinderbox. Eastgate Systems, Watertown MA.

[21] Mark Bernstein and Stee McMorris. 2022. Links of Darkness: Hypertext and Horror. In Proc. Hypertext 2022. ACM, New York, 73–79.

[22] Sven Birkerts. 1994. The Gutenberg Elegies. Faber and Faber, Boston.

[23] Ann Blair. 2010. Too much to know : managing scholarly information before the modern age. Yale University Press, New Haven.

[24] Jay David Bolter. 1991. Writing Space. Lawrence Erlbaum Associates, Hillsdale, NJ.

[25] James Boswell. 1791. Life of Johnson. Henry Baldwin, London.

[26] Tom Bradshaw. 2004. Reading at risk : a survey of literary reading in America. National Endowment for the Arts, Washington, D.C. xiii, 47 p pages.

[27] Richard. Brautigan. 1967. All watched over by machines of loving grace. Communication Company, San Francisco. [28] pages.

[28] Michael Brenner. 1996. 3. A New Learning: The Lehrhaus Movement. In The Renaissance of Jewish Culture in Weimar Germany. Yale University Press, New Haven, 69–99.

[29] Gerard Brett. 1954. The Automata in the Byzantine “Throne of Solomon”. Speculum 29, 3 (1954), 477–487.

[30] F.P. Brooks. 1995. The Mythical Man-Month. Addison-Wesley, Reading, MA.

[31] Rosalind Brown. 2024. Practice. Farrar, Straus and Giroux, New York. 208 pages.

[32] Stephen Budiansky. 2021. Journey to the edge of reason : the life of Kurt Gödel. Oxford University Press, New York. 368 pages.

[33] Andrew M. Burger, Barry D. Meyer, Cindy P. Jung, and Kevin B. Long. 1991. The virtual notebook system. In Proceedings of the third annual ACM conference on Hypertext. ACM, San Antonio, Texas, USA, 395–401.

[34] V. Bush. 1945. As we may think. Atlantic Monthly July 1945 (1945), 101–108.

[35] Michael. Chabon. 1995. Wonder boys. Villard Books, New York. 368 pages.

[36] Michael Chabon. 2008. Trickster in a Suit of Lights. In Maps and legends. McSweeney's Books, San Francisco.

[37] Sara J. Charles. 2024. The Medieval Scriptorium: Making Books in the Middle Ages. Reaktion Books, London. 350 pages.

[38] Andy Clark and David J. Chalmers. 1998. The Extended Mind. Analysis 58, 1 (1998), 7–19.

[39] John Clute. 2014. Stay. Beccon Publications, Essex.

[40] Richard Cockett. 2023. Vienna: How the City of Ideas Created the Modern World. Yale University Press, New Haven. 465 pages.

[41] J. Conklin. 1987. Hypertext: An Introduction and Survey. IEEE Computer 1987, September (1987), 17–41.

[42] William Wallace Cook. 1928. Plotto; a new method of plot suggestion for writers of creative fiction. Ellis publishing company, Battle Creek, Mich. 300 pages.

[43] Robert Coover. 1992. The End of Books. New York Times Book Review 21, 6 (1992), 23–25.

[44] Gregory Crane. 1987. From the old to the new: intergrating hypertext into traditional scholarship. In Proceedings of the ACM conference on Hypertext. ACM, New York, 51–55.

[45] Gregory Crane. 1996. Building a digital library: the Perseus project as a case study in the humanities. In Proceedings of the first ACM international conference on Digital libraries. ACM, New York, 3–10.

[46] Maria C. Dadarlaty, Ryan A. Canfeld, and Amy L. Osborn. 2023. Neural Plasticity in Sensorimotor Brain–Machine Interfaces. Annual Review of Biomedical Engineering 25 (2023), 51–76.

[47] Elanor Dickey. 2017. Classical Scholarship: The Byzantine Contribution. In The Cambridge Intellectual History of Byzantium, Anthony Kaldellis and Niketas Siniossoglou (Eds.). Cambridge University Press, Cambridge, 63–78.

[48] Emily Dickinson and R. W. Franklin. 1998. The poems of Emily Dickinson. Belknap Press of Harvard University Press, Cambridge, Mass. 3 v. pages.

[49] Andreas Dieberger and Jolanda G. Tromp. 1993. The Information City project—a virtual reality user interface for navigation in information spaces Was to appear in. In Proceedings of the Symposium Virtual Reality Vienna (Vienna, Austria) (VRV’93). IDG GMBH, Vienna, Austria.

[50] Alan Dix. 2026. The Abomination of AI. Keynote talk, International Conference on Cognitive Systems and Computer Interaction (ICoSCI 2026), Kuala Terengganu, Malaysia. Accessed 31 July 2026. https://alandix.com/academic/talks/ICOSCI-2026-abomination-of-AI/

[51] Cory Doctorow. 2025. Enshittification : why everything suddenly got worse and what to do about it. MCD, Farrar, Straus and Giroux, New York. 352 pages.

[52] Jane Yellowlees Douglas. 2000. The End of Books – or Books without End?University of Michigan Press, Ann Arbor.

[53] Manish Kumar Dwivedi and Aman Chhonkar. 2025. Writers Guild of America Strike 2023. Emerging Economies Cases Journal 0, 0 (2025), 25166042251397697.

[54] Umberto. Eco. 1995. The search for the perfect language. Blackwell, Oxford, UK; Cambridge, Mass., USA. x, 385 pages.

[55] David Edmonds. 2020. The murder of Professor Schlick : the rise and fall of the Vienna Circle. Princeton University Press, Princeton. 313 pages.

[56] George Eliot. 1856. The Natural History of German Life. Wastminster Review LXVI, 129 (1856), 51–79.

[57] Douglas C. Engelbart. 1962. Augmenting Human Intellect: A Conceptual Framework. Technical Report. Stanford Research Institute. https://www.dougengelbart.org/pubs/augment-3906.html

[58] T. Erickson. 1996. The Design and Long-Term Use of a Personal Electronic Notebook: A Reflective Analysis. In CHI’96. ACM, New York, 11–18.

[59] Ayelet Even-Ezra. 2021. Lines of thought: branching diagrams and the medieval mind. The University of Chicago Press, Chicago. 250 pages.

[60] James Fallows. 1979. Who's a Snob and Who's Not. Washington Monthly June (1979), 10–19.

[61] Ronan Farrow and Andres Marantz. 2026. Moment Of Truth: Can Sam Altman Be trusted?The New Yorker April (2026), 32–51.

[62] Stanley Eugene. Fish. 1980. Is there a text in this class? : The authority of interpretive communities. Harvard University Press, Cambridge, Mass. viii, 394 pages.

[63] Tiago Forte. 2022. Building a second brain : a proven method to organize your digital life and unlock your creative potential. Atria Books, New York. 260 pages.

[64] Foucault, Michel. 1977. What is an Author? In Language, counter-memory, practice : selected essays and interviews. Cornell University Press, Ithaca, N.Y.

[65] Jerry Fowler, Donald G. Baker, Ross Dargahi, Vram Kouramajian, Hillary Gilson, Kevin Brook Long, Cynthia Petermann, and G. Anthony Gorry. 1994. Experience with the virtual notebook system: abstraction in hypertext. In Proceedings of the 1994 ACM Conference on Computer Supported Cooperative Work (Chapel Hill, North Carolina, USA) (CSCW ’94). Association for Computing Machinery, New York, NY, USA, 133–143. https://doi.org/10.1145/192844.192892

[66] Martin Fowler. 1999. Refactoring. Addison-Wesley, Reading, MA.

[67] Hans-Georg Gadamer. 1989. Truth and method. Crossroad, New York. xxxviii, 594 pages.

[68] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, Meng Wang, and Haofen Wang. 2024. Retrieval-Augmented Generation for Large Language Models: A Survey. arXiv (2024), 21.

[69] William Gibson. 1984. Neuromancer. Berkeley Publishing Group, New Yorker.

[70] Peter Godfrey-Smith. 2019. Octopus experience. Animal Sentience 4, 26 (2019), 3.

[71] Anthony Gottlieb. 2025. Ludwig Wittgenstein : philosophy in an age of airplanes. Yale University Press, New Haven. 232 pages.

[72] Clement Greenberg. 1939. Avant-Garde and Kitsch. Partisan Review (1939), 34–49.

[73] G. M. A. Grube. 1965. The Greek and Roman critics. University of Toronto Press, Toronto. xi, 372 pages.

[74] G. Gygax and D. Arneson. 1974. Dungeons & Dragons: Rules for Fantastic Medieval Wargames Campaigns Playable with Paper and Pencil and Miniature Figures. Tactical Studies Rules, Lake Geneva, WI.

[75] F.G. Halasz. 1988. Reflections on Notecards: Seven Issues for the Next Generation of HyperMedia Systems. Commun. ACM 31, 7 (1988), 836–852.

[76] Charlie Hargood, David Millard, Ulrike Sperling, and Alex Mitchell. 2023. The Authoring Problem: An Introduction. In The Authoring Problem, Charlie Hargood, David Millard, Ulrike Sperling, and Alex Mitchell (Eds.). Springer Verlag, Cham, CH, 1–14.

[77] N. Katherine Hayles. 2017. Unthought : the power of the cognitive nonconscious. The University of Chicago Press, Chicago; London. 250 pages.

[78] Jillian M. Hess. 2022. How romantics and victorians organized information : commonplace books, scrapbooks, and albums. Oxford University Press, Oxford. 303 pages.

[79] Michael A. Hiltzik. 1999. Dealers of lightning : Xerox PARC and the dawn of the computer age. HarperBusiness, New York. xxviii, 448 pages.

[80] Richard Holeton. 2001. Figurski At Findhorn on Acid. Eastgate Systems, Inc., Watertown, MA.

[81] Nick Hornby. 2008. Shakespeare Wrote for Money. McSweeney's, San Francisco, CA, USA.

[82] Rose Horowitch. 2024. The Elite College Students Who Can't Read Books. The Atlantic November (November 2024), 14. https://www.theatlantic.com/magazine/archive/2024/11/the-elite-college-students-who-cant-read-books/679945/

[83] Wolfgang. Iser. 1978. The act of reading : a theory of aesthetic response. Johns Hopkins University Press, Baltimore. xii, 239 pages.

[84] Shelley Jackson. 1996. Patchwork Girl: by Mary/Shelley/and Herself. Eastgate Systems, Inc., Watertown MA.

[85] Michael Joyce. 1990. afternoon, a story. Eastgate Systems, Inc., Watertown, MA.

[86] Michael Joyce. 1991. Storyspace As a Hypertext System for Writers and Readers of Varying Ability. In Proceedings of the Third Annual ACM Conference on Hypertext (San Antonio, TX, USA) (HYPERTEXT ’91). ACM, New York, NY, 381–387. https://doi.org/10.1145/122974.125110

[87] Michiko Kakutani. 1997. Never-Ending Saga. The New York Times Magazine September 28, 1997 (1997), 40–41.

[88] Anthony Kaldellis. 2024. The new Roman empire : a history of Byzantium. Oxford University Press, New York, NY. 1133 pages.

[89] Scott Kennedy and Anthony Kaldellis. 2023. Thucydides in Byzantium. In The Cambridge Companion to Thucydides, Polly Low (Ed.). Cambridge University Press, Cambridge, United Kingdom; New York, NY, 248–263.

[90] Sofia Kitromili, James Jordan, and David E. Millard. 2020. What Authors Think about Hypertext Authoring. In Proceedings of the 31st ACM Conference on Hypertext and Social Media. ACM, New York, 9–16.

[91] Benjamín Labatut. 2023. The maniac. Penguin Press, New York. 354 pages.

[92] George P. Landow. 1997. Hypertext 2.0: The Convergence of Contemporary Critical Theory and Technology, 2nd edn. Johns Hopkins Press, Baltimore.

[93] D. H. Lawrence. 1923. Studies in classic American literature. T. Seltzer, New York. ix, 264 pages.

[94] F. R. Leavis. 1948. The great tradition; George Eliot, Henry James, Joseph Conrad. G.W. Stewart, New York. 266 pages.

[95] Douglas B. Lenat, Alan Borning, David McDonald, Craig Taylor, and Steven Weyer. 1983. Knoesphere: Building Expert Systems with Encyclopedic Knowledge Proceedings of the Eighth International Joint Conference on Artificial Intelligence - Volume 1. In Knoesphere: Building Expert Systems with Encyclopedic Knowledge Proceedings of the Eighth International Joint Conference on Artificial Intelligence - Volume 1. Morgan Kaufmann Publishers Inc., Karlsruhe, West Germany San Francisco, CA, USA, 167–169.

[96] Jill Lepore. 2026. We, The Robots. The New Yorker March 30, 2026 (2026), 20–25.

[97] Michael Lesk. 1989. What to do when there's too much information. In Proceedings of the Second Annual ACM Conference on Hypertext (Pittsburgh, Pennsylvania, USA) (HYPERTEXT ’89). Association for Computing Machinery, New York, NY, USA, 305–318. https://doi.org/10.1145/74224.74249

[98] Lewis-Kraus. 2026. I Claudius. The New Yorker February 16 & 23, 2026 (2026), 52–65.

[99] Jack Lindsey. 2025. Emergent Introspective Awareness in Large Language Models. Technical Report. Anthropic. https://arxiv.org/abs/2601.01828

[100] N. J. Lowe. 2000. The classical plot and the invention of Western narrative. Cambridge University Press, Cambridge; New York. xiii, 293 p. pages.

[101] Norman Macrae. 1999. John von Neumann : the scientific genius who pioneered the modern computer, game theory, nuclear deterrence, and much more. American Mathematical Society, [Providence, RI]. x, 405 pages.

[102] Catherine C. Marshall. 1998. Toward an Ecology of Hypertext Annotation. In Proceedings of the Ninth ACM Conference on Hypertext and Hypermedia : Links, Objects, Time and Space—structure in Hypermedia Systems (Pittsburgh, PA, USA) (HYPERTEXT ’98). ACM, New York, NY, USA, 40–49. https://doi.org/10.1145/276627.276632

[103] Catherine C. Marshall. 2009. Reading and Writing the Electronic Book. Morgan & Claypool Publishers, San Rafael, CA, USA. 200 pages.

[104] Robert C. Martin. 2009. Clean code : a handbook of agile software craftsmanship. Prentice Hall, Upper Saddle River, NJ. xxix, 431 pages.

[105] Adrienne Mayor. 2018. Gods and robots : myths, machines, and ancient dreams of technology. Princeton University Press, Princeton. 275 pages.

[106] R. W. McCutcheon. 2015. Silent Reading in Antiquity and the Future History of the Book. Book History 18 (2015), 1–32.

[107] Eden Medina. 2011. Cybernetic revolutionaries : technology and politics in Allende's Chile. MIT Press, Cambridge, Mass. xiv, 326 pages.

[108] James R. Meehan. 1977. TALE-SPIN, an interactive program that writes stories. In TALE-SPIN, an interactive program that writes stories.Morgan Kaufmann Publishers Inc., Cambridge, USA San Francisco, CA, USA, 91–98.

[109] Louis Menand. 2021. The free world :art and thought in the Cold War. Farrar, Straus and Giroux, New York. 880 pages.

[110] Laura Miller. 1998. www.claptrap.com. New York Times Book Review (1998), 43.

[111] Arnaldo. Momigliano. 1990. The classical foundations of modern historiography. Vol. v. 54. University of California Press, Berkeley. xiv, 162 pages.

[112] Neville Morley. 2026. Thucydiocies: How Not to Read Thucydides in Contemporary Political Discourse. Public Humanities 2 (2026), e40.

[113] Jason Morningstar. 2026. The Sickle Hour. Bully Pulpit Games LLC. 18 pages.

[114] Stuart Moulthrop. 1989. Hypertext and ‘the Hyperreal’. In Hypertext and ‘the Hyperreal’. ACM, Pittsburgh, 291–298.

[115] Stuart Moulthrop. 2026. Parrots on a Wet, Black Bough: Facing into AI Art. Electronic Book Review (2026).

[116] Stuart Moulthrop and Dene Grigar. 2017. Traversals : the use of preservation for early electronic writing. The MIT Press, Cambridge, MA. 273 pages.

[117] Jocelyne Nanard and Marc Nanard. 1993. Should Anchors Be Typed Too? An Experiment with MacWeb. In Proceedings of the Fifth ACM Conference on Hypertext (Seattle, WA, USA) (HYPERTEXT ’93). ACM, New York, NY, USA, 51–62. https://doi.org/10.1145/168750.168767

[118] T. Nelson. 1982. Literary Machines. Mindscape Press, 1987.

[119] Jakob Nielsen. 1989. The Matters That Really Matter for Hypertext Usability. In Proceedings of the Second Annual ACM Conference on Hypertext (Pittsburgh, PA, USA) (HYPERTEXT ’89). Association for Computing Machinery, New York, NY, USA, 239–248. https://doi.org/10.1145/74224.74244

[120] T. Oren, G. Solomon, K. Kreitman, and A. Don. 1990. Guides: Characterizing the Interface. In The Art of Human Computer Interface Design, B. Laurel (Ed.). Addison Wesley, Reading, MA, USA, 367–383.

[121] Stephen Orgel. 2015. The reader in the book : a study of spaces and traces. Oxford University Press, Oxford, United Kingdom. 171 pages.

[122] Paul Otlet. 1934. Traité de documentation : le livre sur le livre, theéorie et pratique. Editiones Mundaneum, Bruxelles. 431, [12], viii pages.

[123] Susanne Paulus. 2023. Back to School in Babylonia. Institute for the Study of Ancient Cultures, Chicago.

[124] Luke V. Pitcher. 2023. Thucydides in Greek and Roman Historiography. In The Cambridge Companion to Thucydides, Polly Low (Ed.). Cambridge University Press, Cambridge, United Kingdom; New York, NY, 292–247.

[125] Philip Pullman and Simon Mason. 2017. Dæmon voices : essays on storytelling. David Fickling Books, Oxford, UK. 480 pages.

[126] B. Reeves and C. Nass. 1996. The Media Equation - How People Treat Computers, Television, and New Media Like Real People and Places. Cambridge University Press, Cambridge, MA.

[127] Eleanor Robson. 2011. The Production and Dissemination of Scholarly Knowledge. In The Oxford Handbook of Cuneiform Culture, Radner Karen and Robson Eleanor (Eds.). Oxford University Press, Oxford, UK, 557–577.

[128] James Romm. 2021. What Happened At Masada?The New York Review of Books January 14, 2021 (2021), 4.

[129] Matthew Schwartz. 2026. Vibe physics: The AI grad student. Retrieved April 14, 2026 from https://www.anthropic.com/research/vibe-physics

[130] Bernard Shaw. 1905. The author's apology from Mrs. Warren's profession. Brentano's, New York. 66 pages.

[131] Ben Shneiderman. 1989. Reflections on Authoring, Editing, and Managing Hypertext. In The Society of Text, Edward Barrett (Ed.). MIT Press, Cambridge, MA, 115–131.

[132] B. Shneiderman. 1993. A nonanthropomorphic style guide: Overcoming the Humpty Dumpty syndrome. In Sparks of Innovation in Human-Computer Interaction. Greenwood Publishing Group, Westwood, CT, 331–341.

[133] B. Shneiderman and M. Muller. 2023. On AI anthropomorphism. Human-Centered AI (2023).

[134] Karl Sigmund. 2017. Exact Thinking in Demented Times: The Vienna Circle and the Epic Quest for the Foundations of Science. Basic Books, New York, NY. 491 pages.

[135] Jane Smiley. 1996. Say It Ain't So, Huck. Harper's Magazine (1996), 61–67.

[136] Susan Sontag. 1964. Against Interpretation. Evergreen Review 34 (1964), 5.

[137] Lars. Spuybroek. 2011. Sympathy of things : Ruskin and the ecology of design. V2 Publishing : NAi Publishing Available in North, South and Central America through D.A.P./Distributed Art Publishers Inc., [Rotterdam] New York, NY. 399 pages.

[138] Wallace Stevens. 1937. The man with the blue guitar & other poems. Alfred A. Knopf, New York. 82 pages.

[139] Lyndsey Stonebridge. 2024. We are free to change the world : Hannah Arendt's lessons in love and disobedience. Hogarth, London; New York. 346 pages.

[140] P. David Stotts and Richard Furuta. 1989. Petri-net based hypertext: Document structure with browsing semantics. ACM Transactions on Office Information Systems 7, 1 (1989), 3–29.

[141] Dennis Yi Tenen. 2024. Literary Theory for Robots: How Computers Learned to Write. W. W. Norton & Company, New York, NY, USA. 176 pages.

[142] Thucydides. 1998. History of the Peloponnesian War.

[143] Thucydides. and Jeremy. Mynott. 2013. The war of the Peloponnesians and the Athenians. Cambridge University Press, Cambridge; New York. lxiv, 690 pages.

[144] James Turner. 2015. Philology: The Forgotten Origins of the Modern Humanities. Princeton University Press, Princeton, NJ, USA. 574 pages.

[145] Ron K. Unz. 1985. The Surplus of the Athenian Phoros. Greek, Roman, and Byzantine Studies 26, 1 (1985), 21–42.

[146] Shannon Vallor. 2024. The AI mirror : how to reclaim our humanity in the age of machine thinking. Oxford University Press, New York, NY. 257 pages.

[147] H. Van Dyke Parunak. 1991. Don't Link Me in: Set Based Hypermedia for Taxonomic Reasoning. In Proceedings of the Third Annual ACM Conference on Hypertext (San Antonio, TX, USA) (HYPERTEXT ’91). ACM, New York, NY, USA, 233–242. https://doi.org/10.1145/122974.122998

[148] Jill Walker. 1999. Piecing together and tearing apart: finding the story in afternoon. In Proceedings of the Tenth ACM Conference on Hypertext and Hypermedia: Returning to Our Diverse Roots: Returning to Our Diverse Roots (Darmstadt, Germany) (HYPERTEXT ’99). Association for Computing Machinery, New York, NY, USA, 111–117. https://doi.org/10.1145/294469.294496

[149] Jo. Walton. 2003. Tooth and claw. Tor, New York. 253 pages.

[150] Joseph. Weizenbaum. 1976. Computer power and human reason : from judgment to calculation. W. H. Freeman, San Francisco. xii, 300 pages.

[151] Stephen Wolfram. 2025. What's Really Going On in Machine Learning? Some Minimal Models. Wolfram Media, Champaign, IL, USA. 107 pages.

[152] Polle T. Zellweger, Anne Mangen, and Paula Newman. 2002. Reading and Writing Fluid Hypertexts. In Reading and Writing Fluid Hypertexts, Stuart Moulthrop, Ken Anderson, and Jamie Blustein (Eds.). ACM, College Park, Maryland, 45–54.

[153] Siyi Zhu and Joel Chan. 2023. Exploring Distributed Synthesis: In-Progress Findings from Guided Tours of Scholarly Knowledge Synthesis Work Practices with A Distributed Lens. In Companion Publication of the 2023 Conference on Computer Supported Cooperative Work and Social Computing. ACM, New York, 328–332.