Lisa Gitelman

Foreword

By Edo Segal

The thing I should have noticed first was the format.

Not the technology. Not what Claude could do or how fast it could do it. The format. The fact that I poured a conversational, iterative, genuinely strange production process into the shape of a printed book and never once asked whether the shape was distorting the thing I was trying to say.

I described the collaboration honestly. I disclosed Claude's role in the Foreword, dedicated a chapter to the writing process, included Claude's own reflections unedited. I thought I was being transparent. I was. But transparency about the contents does not help you if the container itself is doing the hiding.

The linear chapter structure implied I wrote this book the way books get written — one idea building on the next, in sequence, from premise to conclusion. That is not what happened. What happened was a dialogue. Ideas arrived out of order, got rearranged, got discarded, got resurrected by a connection Claude surfaced three conversations later. The book you hold tells a story of sequential development. The process was anything but.

I did not see this until I encountered Lisa Gitelman.

Gitelman is a media historian who studies something most of us never think about: the conventions that surround a technology and determine what it means. Not what it can do — what it means. The protocols, she calls them. The social practices, legal frameworks, economic models, and habits of use that gather around a technological nucleus and define its cultural role. Her signature insight is that these protocols are constructed, not natural. They feel inevitable once they settle. They are not. They are the residue of institutional negotiations, power dynamics, and accumulated decisions made during the brief window when everything is still in play.

That window is now. For AI.

The conventions that will govern how AI-assisted work is credited, evaluated, priced, and understood are being formed right now — not through philosophy but through the accumulated weight of decisions made by technology companies, publishers, regulators, and practitioners like me. Every choice I made about how to present this book — the attribution on the title page, the inclusion of Claude's reflections, the disclosure in the Foreword — was a proposal for a convention that does not yet exist. Whether those proposals survive depends on forces far larger than one book.

Gitelman gave me the vocabulary to see what the format was concealing. This volume explores what that vocabulary reveals about AI, authorship, and the invisible architecture shaping everything we build in this unsettled moment.

— Edo Segal ^ Opus 4.6

About Lisa Gitelman

1956-present

Lisa Gitelman (born 1956) is an American media historian and cultural theorist, Professor of English and Media, Culture, and Communication at New York University. Trained in American Studies, she has devoted her career to understanding how media technologies acquire cultural meaning through the institutional protocols that surround them. Her book Always Already New: Media, History, and the Data of Culture (2006) argued that media are never simply technologies but are always embedded in social practices, legal frameworks, and habits of use that determine their cultural role — conventions that feel natural once settled but are always historically constructed. She edited the influential collection "Raw Data" Is an Oxymoron (2013), which demonstrated that data is never a neutral given but is always shaped by the instruments, institutions, and assumptions that produce it. In Paper Knowledge: Toward a Media History of Documents (2014), she examined how the material properties and institutional contexts of documents shape what counts as knowledge. Co-founder of NYU's Digital Theory Lab, which has studied deep learning since 2018, Gitelman has more recently turned her attention to AI-generated media, delivering lectures on "typographical hallucinations" in image-generating systems. Her work is foundational to the fields of media archaeology and critical data studies, providing essential tools for understanding how the conventions forming around AI will shape its cultural role for generations.

Chapter 1: Every New Medium Begins in Someone Else's Clothes

Film was called moving pictures. Radio was called wireless telegraphy. Television was called visual radio. The phonograph was called a talking machine. In each case, a technology that would eventually reshape the conditions of cultural life entered the world without a vocabulary of its own, reaching instead for the nearest available language — the categories, conventions, and evaluative frameworks of the medium it was displacing. The borrowed language fit poorly. It described what the new medium looked like from the vantage point of the old one, which is to say it described almost nothing about what the new medium would become. But the borrowed language was all there was, because the cultural tailoring of a new medium — the slow, contested, institutionally mediated process by which a medium acquires conventions adequate to its own operations — had not yet begun.

Lisa Gitelman's career has been devoted to understanding this process with a precision that most media commentary lacks. In Always Already New, she demonstrated that media are never simply technologies. They are technologies embedded in what she calls "protocols" — "a vast clutter of normative rules and default conditions, which gather and adhere like a nebulous array around a technological nucleus." The protocols include social practices, institutional frameworks, legal regimes, economic models, habits of use, and the tacit assumptions that users bring to a medium so reflexively that the assumptions become invisible. The technology is the nucleus. The protocols are the culture that forms around it. And the two cannot be separated, because what a medium is — not what it can do in a laboratory, but what it does in the world — is determined by the interaction between its technical capabilities and the protocols that govern their use.

The protocols of a new medium are never invented from scratch. They are borrowed from existing media and then, over time, modified, contested, and eventually replaced by protocols specific to the new medium. This borrowing is not a failure of imagination. It is a structural necessity. The users of a new medium need some framework for understanding what the medium is and what it is for, and the only frameworks available are those provided by existing media. The earliest radio broadcasters structured their programs as lectures and concerts because the protocols of the lecture hall and the concert hall were the available models. The earliest filmmakers framed their shots as theatrical tableaux because the protocols of the stage were what they knew. The borrowed protocols shaped not only what the new medium was used for but what it was understood to be — and the understanding, in turn, constrained what users could imagine doing with it.

Artificial intelligence, as a medium of cultural production, is currently wearing borrowed clothes. The protocols through which AI-assisted texts are produced, circulated, and evaluated are protocols borrowed from print culture — the medium whose clothes fit most readily, because AI-assisted production generates the same kind of artifact that print culture generates: written text organized into documents. The borrowing is so thorough that AI-assisted texts look, at first glance, indistinguishable from traditionally authored texts. They have title pages and tables of contents. They have named authors. They have chapters that develop sequential arguments. They bear the paratextual apparatus of the printed book — acknowledgments, source lists, author biographies — that has served for centuries as the institutional scaffolding of individual authorship.

Edo Segal's The Orange Pill is a case in point. The book is credited to "Edo Segal ^ Opus 4.6." It contains twenty chapters organized into five parts, a foreword, a prologue, acknowledgments, and a list of selected sources. It is, in every visible respect, a book — a printed or printable codex-format artifact that conforms to the protocols of single-authored nonfiction. A reader encountering it without context would find nothing in its physical or digital format to suggest that it was produced through any process other than the one the format implies: a single author, working through sustained private effort, translating individual thought into sequential prose.

But the format is borrowed. The process that produced the text was not the process that the format implies. Segal describes a production method that was conversational, iterative, and distributed across the boundary between a human intelligence and a machine. He describes sessions in which he offered half-formed ideas to Claude and received structural scaffolding in return. He describes moments when Claude identified connections between ideas from different chapters that Segal had not seen — connections that, by his own account, "belonged to the collaboration" rather than to either participant alone. He describes a production timeline — a 187-page draft written on a transatlantic flight — that is incompatible with the conventions of the solitary author laboring over months or years in private.

The format of the printed book conceals all of this. Not deliberately, not through any act of deception, but structurally — because the format was designed for a different process, and when a different process is poured into it, the format smooths the difference away. The linear chapter structure implies a linear process of composition. The single-author attribution implies a single source of creative agency. The sequential argument implies a coherent intellectual development from premise to conclusion. None of these implications accurately represent what Segal describes. The borrowed clothes conceal the body they are dressing.

Gitelman's framework makes this concealment visible — not as a scandal to be exposed but as a structural feature of every media transition. The protocols of the old medium always shape the early products of the new one, and the shaping always involves a kind of distortion: the new medium's distinctive features are compressed into the old medium's categories, and the compression hides what does not fit. The earliest films hid the potential of montage because the protocols of theatrical staging did not have a category for it. The earliest radio broadcasts hid the potential of the disc jockey and the call-in show because the protocols of the lecture hall did not imagine audience participation. The earliest AI-assisted texts hide the distinctive features of conversational, distributed, human-machine production because the protocols of individual authorship do not have categories for them.

The distortion is not permanent. It belongs to what Gitelman identifies as the unsettled period — the phase of a medium's development when its protocols are still forming, when the borrowed clothes have not yet been replaced by clothes of the medium's own. The unsettled period is characterized by visible awkwardness, by the strain of a body that does not fit the garments it is wearing, by the anxiety of participants who sense that the rules they are following are inadequate but who do not yet have access to better ones. The awkwardness is productive. It is the force that drives the development of new protocols — protocols adequate to the new medium's distinctive operations, protocols that can accommodate what the borrowed ones cannot.

The Orange Pill exhibits this productive awkwardness throughout. The notation "Edo Segal ^ Opus 4.6" on the title page is a small but revealing case. The caret symbol has no conventional meaning in the context of authorial attribution. It is borrowed from mathematics and programming, domains where it signifies exponentiation or bitwise operations — neither of which maps onto the relationship between a human author and an AI collaborator. The notation is an improvisation, a one-off solution to a problem that the existing protocols of authorship do not address: how to credit a non-human participant in a process that the concept of authorship was designed to describe as exclusively human. The improvisation is visible as an improvisation. It does not look natural or inevitable. It looks like what it is — a placeholder, a gesture toward a convention that does not yet exist, a borrowed symbol repurposed for a situation that no existing symbol was designed to fit.

The improvisation extends to the book's treatment of Claude's contributions. Segal includes two sections explicitly attributed to Claude — a "Reflection Before the First Word" and a "Reflection After the Last Word" — presented as Claude's own prose, written by the machine and included without editorial modification. The inclusion has no precedent in the conventions of print culture. The closest analogies — a translator's preface, an editor's introduction — are inadequate because they involve human collaborators whose contributions are understood within existing frameworks of intellectual partnership. Claude's reflections are something else. They are machine-generated text, included in a human-authored book, attributed to the machine, and presented as a form of testimony about the production process. The protocols of print culture have no category for this kind of inclusion, and its presence in the text produces the specific discomfort of a convention that has not yet formed — the awareness that something is happening for which the existing rules provide no guidance.

Gitelman's concept of protocols illuminates why the borrowed-clothes phase matters far more than its temporary awkwardness might suggest. Protocols are not merely conventions. They are the institutional infrastructure through which a medium's cultural role is defined and maintained. The protocols of print culture — copyright law, editorial review, the attribution of texts to named individuals, the evaluation of texts by criteria that assume individual production — are not incidental features of the publishing industry. They are the mechanisms through which the cultural value of written work is assessed, the economic rewards of authorship are distributed, and the responsibility for textual claims is assigned. When these protocols are borrowed and applied to a medium they were not designed for, the misfit has consequences that extend far beyond aesthetics. It affects who gets credit, who gets paid, who gets blamed, and how the cultural significance of the new medium's products is understood.

The consequences are already visible. Copyright law in most jurisdictions requires a human author for a work to be copyrightable. The US Copyright Office has determined that AI-generated content cannot receive copyright protection, but that works containing AI-generated elements may be copyrightable if a human exercised "sufficient creative control." The phrase "sufficient creative control" is a protocol in formation — a legal concept with no established meaning, no precedent, and no settled test. It is the legal system's version of the caret symbol on Segal's title page: an improvisation, a placeholder, a borrowed concept pressed into service for a situation it was not designed to address. The legal system is wearing borrowed clothes too, and the fit is no better.

The history of media transitions suggests that the borrowed-clothes phase is not merely an inconvenience to be endured. It is the phase in which the most consequential decisions about the new medium are made. The conventions that form during the unsettled period tend to persist long after the period has ended. They become the invisible infrastructure of the medium — the assumptions so deeply embedded in practice that they cease to be visible as assumptions. The conventions of individual print authorship, for example — the ideas that a text has a single author, that the author owns the text, that the text expresses the author's thought — were constructed during print culture's own unsettled period, through negotiations involving publishers, courts, booksellers, and the emerging concept of intellectual property. Once constructed, they became so thoroughly naturalized that considerable intellectual effort was required to see them as conventions at all rather than as descriptions of what a book simply is.

The conventions now forming around AI-assisted cultural production will be similarly consequential. How credit is attributed, how quality is evaluated, how the relationship between human and machine contributions is understood — these matters are being decided now, through the accumulated weight of individual decisions made by early practitioners, institutional gatekeepers, and market forces. Segal's choices about disclosure, attribution, and the inclusion of Claude's reflections are not merely personal decisions about how to present one book. They are proposals — concrete, enacted proposals for conventions that do not yet exist. Whether they are adopted, modified, or rejected by subsequent practitioners and institutions will depend on forces that extend far beyond any single book.

The medium is always already new. By the time it has been recognized as a medium, it has already begun developing the protocols that will shape its cultural role for generations. The clothes are being tailored now. Every stitch matters — not because any individual stitch is irreversible, but because the accumulated weight of stitches produces a garment that, once worn, becomes difficult to take off.

---

Chapter 2: The Moving-Pictures Stage

When the Lumière brothers screened their first films in the Grand Café in Paris on December 28, 1895, the audience saw what the existing categories allowed them to see: photographs that moved. Workers leaving a factory. A train arriving at a station. A baby being fed. A gardener being sprayed with his own hose. The films were exhibited as novelties, attractions whose primary appeal was the sheer mechanical wonder of captured motion projected onto a wall. The criteria for evaluating them were borrowed from photography — clarity, composition, fidelity to the visual world — because no criteria specific to the new medium existed. If a moving picture was sharp and well-framed, it was a good moving picture. If it was blurry or poorly composed, it was not. The evaluation was coherent within its borrowed framework and utterly inadequate to what film would become.

No one in the Grand Café imagined narrative cinema. No one imagined the close-up, the tracking shot, the dissolve, the montage sequence, the continuity editing system that would eventually make spatial and temporal relationships legible to audiences through visual grammar alone. The medium's potential was invisible because it was being perceived through categories that could not accommodate it. The categories were not wrong — they accurately described what film was at that moment. They were profoundly misleading about what film could be, because what film could be was not deducible from what film was. It could only be discovered through practice, through formal experiments conducted by practitioners who did not know what they were building, through the slow accumulation of techniques that would, over decades, constitute a medium no longer beholden to the protocols of photography or theater.

Lisa Gitelman's historical work on media transitions reveals that this pattern — the initial use of a new medium to perform faster or cheaper versions of what existing media already do, followed by the gradual discovery of capabilities specific to the new medium — is structural rather than anecdotal. It recurs with such consistency across the history of communication technologies that it constitutes something approaching a law of media development. The phonograph was initially marketed as a dictation device — a way to do faster what handwriting already did. Its transformation into a music-playback technology required not just technical refinement but the wholesale development of new protocols: recording industry economics, performance contracts, retail distribution, the cultural category of the "record" as a unit of musical experience. Radio was initially deployed as point-to-point communication — a wireless telegraph — and its transformation into a broadcast medium required the invention of programming formats, advertising models, regulatory frameworks, and the cultural concept of the "audience" as a dispersed, simultaneous public. In each case, the medium's distinctive capabilities were invisible during the borrowed-clothes phase and became visible only through the development of protocols specific to the new medium.

AI-assisted cultural production is at the moving-pictures stage. The parallel is precise enough to be analytically useful. The current deployment of AI tools — writing code faster, drafting documents more efficiently, generating images on demand — corresponds to the Lumière brothers' use of the camera to record workers leaving a factory. The technology is being used to do faster versions of what existing media already do. The evaluation criteria are borrowed from the protocols of the media being displaced: AI-assisted text is judged by whether it reads like good individually authored text, AI-assisted code is judged by whether it functions like good individually written code, AI-assisted images are judged by whether they look like good individually created images. The borrowed criteria are coherent. They are also, if the historical pattern holds, profoundly inadequate to what AI-assisted production will become.

The Orange Pill occupies an interesting position within this parallel. It is, in most respects, a moving-pictures artifact — a book that uses AI to produce a faster, smoother version of what books have always been. The structure is conventional. The argument is sequential. The voice belongs to a named individual. The protocols of print culture govern every visible element of its presentation. But the text also contains moments that point beyond the moving-pictures stage, moments where something specific to the new medium becomes fleetingly visible before disappearing back into the borrowed format.

The most revealing of these moments is Segal's description of the laparoscopic surgery connection. He was stuck — unable to find the analytical pivot between acknowledging the philosopher Byung-Chul Han's critique of frictionlessness and demonstrating that friction does not disappear when mechanical difficulty is removed but instead ascends to a higher cognitive level. He described the impasse to Claude. Claude responded with an example from surgical history: when laparoscopic techniques replaced open surgery, surgeons lost the tactile friction of hands inside the body but gained the ability to perform operations that open hands could never attempt. The friction did not vanish. It relocated upward.

Segal writes: "Neither of us owns that insight. The collaboration does." The statement is a break from the protocols of individual authorship, and it is also a glimpse of something specific to the new medium — a mode of intellectual production that is genuinely collaborative in a way that existing categories of collaboration do not capture. The insight did not exist in Segal's thinking before the exchange. It did not exist in Claude's training data as a pre-formed connection between Han's philosophy and surgical history. It emerged from the collision between a specific question and a specific associative response, mediated by a conversational interface that allowed the collision to occur in real time and to be immediately incorporated into the developing argument. This is not a faster version of what books have always done. It is something else — a mode of intellectual production whose distinctive features are visible only in moments and that the borrowed format of the book cannot fully accommodate.

Gitelman's analysis of early phonograph culture is instructive here. When Thomas Edison first demonstrated the phonograph in 1877, he imagined it primarily as an office technology — a device for recording and playing back dictation. His famous list of anticipated uses placed "letter writing and all kinds of dictation" at the top and "music" near the bottom. The phonograph's eventual cultural role — as a medium of recorded music that would reshape entertainment, industry, and the phenomenology of listening — was invisible to its inventor because the protocols he brought to the device were the protocols of the office, not the concert hall. The device's technical capabilities included the capacity to record and reproduce music, but the cultural significance of that capacity could not be perceived through the evaluative framework of office productivity.

The same gap between technical capability and cultural significance characterizes the current moment of AI-assisted production. The technical capability includes modes of intellectual production that have no precedent in print culture — the real-time, conversational development of ideas through human-machine dialogue, the identification of cross-domain connections across scales of knowledge that no individual mind can hold, the iterative refinement of arguments through rapid feedback loops that compress what would have been months of drafting into hours. These capabilities are technically present. Their cultural significance cannot yet be perceived, because the protocols through which they are evaluated — the protocols of individual authorship, original creation, and scholarly rigor — were designed for a different medium and cannot accommodate what they do not recognize.

Gitelman's work-in-progress on what she calls "typographical hallucinations" — the study of how image-generating AI systems like DALL-E struggle with typographic forms, producing mangled letterforms even as they generate plausible visual contexts — offers an unexpectedly precise lens for understanding this gap. In her lectures at the University of Pennsylvania and the University of Virginia in 2024 and 2025, Gitelman approached AI's typographic failures not as errors to be fixed but as evidence of what the system does and does not "know" about human culture. She asked: "Is there something that DALL-E 3 'knows' about typography, in short, that we don't?" The question is characteristic of Gitelman's method — treating the machine's output not as a success or failure relative to human standards but as a document that reveals the specific kind of knowledge that statistical pattern-matching produces.

The typographical hallucination is a small-scale instance of a large-scale phenomenon. AI systems produce outputs that are plausible within their own statistical framework but that fail when measured against the culturally embedded knowledge that human practitioners bring to the same domain. A typographer knows that letterforms obey specific geometric and historical constraints. DALL-E knows that letterforms appear in specific visual contexts — on signs, in books, on screens — but does not know the constraints that make a given letterform correct or incorrect. The result is text that looks like text from a distance but dissolves into incoherence upon close inspection. The format is right. The content fails.

This is the same structure that Segal identifies in The Orange Pill when he describes Claude producing a confident, well-organized passage connecting Csikszentmihalyi's flow state to a concept attributed to Gilles Deleuze. The passage was rhetorically effective. It sounded like insight. The philosophical reference was wrong in a way that would have been obvious to anyone who had read Deleuze. The format — polished academic prose, confident assertion, seamless integration into the argument — was right. The content failed. And the format concealed the failure, because the conventions of print culture associate the format of confident, well-organized prose with the process of careful scholarship. The association was built for a medium in which the format and the process were reliably linked. In the new medium, they are not.

The moving-pictures stage will end. It always does. Practitioners will discover capabilities specific to the new medium, capabilities that the borrowed protocols cannot accommodate and that will require the development of new ones. The discovery will not come from theoretical analysis. It will come from practice — from experiments conducted by people who do not know what they are building, whose formal innovations will accumulate into conventions that make the medium's distinctive operations legible and evaluable. Gitelman's historical work on the phonograph, the telegraph, and other communication technologies demonstrates that this process is never fast, never smooth, and never directed by any single participant. It is a collective, institutional, market-mediated process in which proposed conventions are tested, adopted, modified, or rejected through the accumulated weight of decisions made by practitioners, gatekeepers, and audiences operating without the guidance of established norms.

The question is not whether AI-assisted cultural production will move beyond the moving-pictures stage. The question is what conventions will emerge to govern it when it does — and who will shape those conventions during the unsettled period when everything is still in play.

---

Chapter 3: Raw Data, Cooked Output

In 2013, Lisa Gitelman edited a collection of essays under a title that has become one of the most cited phrases in contemporary data studies: "Raw Data" Is an Oxymoron. The book's argument was deceptively simple. Data, Gitelman and her contributors demonstrated, is never raw. It is always cooked — always shaped by the instruments that collect it, the categories that organize it, the institutions that commission it, and the assumptions that determine what counts as data in the first place. The word "data" comes from the Latin dare, to give, implying that data is given by the world, found rather than made, a natural resource waiting to be harvested. Gitelman showed that this etymology is a lie embedded in the language. Data is not given. It is taken — extracted from the world through specific instruments, according to specific protocols, for specific purposes, by specific institutions with specific interests. What looks like a neutral description of the world is always already an interpretation, and the interpretive framework is usually invisible because it is embedded in the instruments and institutions that produce the data rather than in the data itself.

The argument has acquired a second life in the age of artificial intelligence, where it illuminates problems that Gitelman could not have anticipated when the book was published — though her theoretical apparatus was already equipped to handle them. The phrase "AI-generated content" performs the same sleight of hand as the phrase "raw data." Both suggest a product that arrives without mediation — content generated by a machine, data given by the world. Both are, upon examination, deeply misleading. AI-generated content is not generated from nothing. It is generated from training data, which was generated from documents, which were generated within institutional contexts governed by specific protocols of formatting, selection, and preservation. The training corpus is not a neutral sample of human knowledge. It is a specific collection shaped by what was digitized, what was publicly available, what was written in English, what survived the filters of platform terms of service, copyright law, and web architecture. When the output is attributed to the machine's generative capacity, the massive human infrastructure of inscription, formatting, and curation that made the output possible disappears from view.

Gitelman's framework demands that this infrastructure be made visible — not as an accusation but as a condition of understanding what the medium actually does. When Claude produces a passage of polished analytical prose, the passage is not generated from a void. It is generated from patterns identified in a corpus that has specific, historically contingent characteristics. The corpus overrepresents certain languages, certain cultural traditions, certain institutional registers. It underrepresents oral traditions, non-Western epistemologies, knowledge that was never written down or that was written down but never digitized. The output inherits these biases not as bugs to be fixed but as structural features of the data from which it was produced. The data was never raw. The output is never neutral. And the format of the output — confident, well-organized, grammatically impeccable prose — conceals the specificity of the data behind an appearance of authoritative generality.

The Orange Pill is produced from this infrastructure and is partially aware of the fact. Segal describes Claude's training data as "the entire history of human thought" — a phrase that reveals the very mystification Gitelman's framework is designed to expose. The training data is not the entirety of human thought. It is a subset of human thought that was written in certain languages, published through certain channels, digitized by certain institutions, and made available under certain licensing regimes. The gap between "the entire history of human thought" and the actual contents of the training corpus is not a minor discrepancy. It is a structural feature of the medium that shapes every output the medium produces. When Claude draws a connection between Csikszentmihalyi's flow psychology and Han's critique of the achievement society, the connection is drawn not from the entirety of human knowledge but from the specific subset of human knowledge that was available, in digital form, in the languages and formats that the training process could ingest. What lies outside that subset — oral traditions, untranslated scholarship, knowledge embedded in practices that were never textualized — is invisible to the system and therefore invisible in the output.

Gitelman's concept of the document is essential here. In Paper Knowledge: Toward a Media History of Documents (2014), she argued that documents are not merely containers for information. They are "epistemic objects" — artifacts that participate in the production of knowledge through their specific material properties, their institutional contexts, and the protocols that govern their creation and circulation. A birth certificate is not a record of a birth. It is a document that constitutes a legal identity through institutional protocols of certification, filing, and retrieval. A scientific paper is not a report of research. It is a document whose format — abstract, methods, results, discussion — shapes what can be claimed and how the claim is evaluated. The format is not neutral. It is constitutive. It determines what counts as knowledge within the institutional framework that the format serves.

AI-generated text is a document in Gitelman's sense. It has a format — the format of confident, well-organized prose — and the format carries epistemological implications. The format of polished academic prose implies that the claims it contains have been produced through processes of research, analysis, and verification. The implication is inherited from print culture, where the format and the process were reliably linked. A passage that reads like careful scholarship was, in print culture, usually produced through careful scholarship. The format functioned as a proxy for the process, and the proxy was generally reliable. In AI-assisted production, the link between format and process is broken. The format persists — the prose is polished, the assertions are confident, the structure is coherent — but the process that produced it is statistical pattern-matching, not research, analysis, or verification. The format implies a depth that the process does not provide.

Segal identifies this problem in The Orange Pill as "the confidence problem" — the gap between Claude's fluent output and the reliability of the claims that fluency implies. The identification is accurate. What Gitelman's framework adds is the recognition that this is not a bug in the technology. It is a feature of the borrowed format. The format of confident prose was developed for a medium in which confidence was earned through process. When the format is borrowed by a medium in which confidence is produced by statistical regularity, the format becomes misleading — not because it is deliberately deceptive, but because the conventions that associated the format with reliability were built for a different medium and have been imported without adjustment.

The Deleuze error that Segal describes is a precise illustration. Claude produced a passage connecting Csikszentmihalyi's flow state to a concept attributed to Deleuze. The passage was confident, well-organized, and rhetorically effective. It read like the product of a scholar who had studied both thinkers and identified a genuine intellectual connection. The philosophical reference was wrong. The format — the document format, in Gitelman's sense — concealed the error by presenting it in the register of scholarly authority. Segal caught the error the next morning, but he notes that he initially accepted it, read it twice, and liked it. The format did what formats do: it shaped the reader's expectations about the reliability of the content, and the expectations were wrong because the format's epistemic guarantees no longer applied.

The implications extend beyond individual errors to the epistemological infrastructure of AI-assisted knowledge production. If the format of AI-generated text carries implicit guarantees of reliability that the production process does not warrant, then every AI-generated claim occupies an uncertain epistemic position — not because the claim is necessarily wrong, but because the reader has no way of knowing, from the format alone, whether the claim was produced through processes that justify confidence or through pattern-matching that produces confidence without justification. The format is the same either way. The difference is invisible.

Gitelman's work with the NYU Digital Theory Lab, which she co-founded and which has studied deep learning since 2018, positions her to recognize this problem with particular clarity. Her 2024-2025 lecture series on "typographical hallucinations" — the study of how image-generating AI systems produce mangled letterforms in otherwise plausible visual contexts — extends the oxymoron framework into the domain of generative AI. The hallucinated letterform is a format that fails: it has the visual appearance of typography without the cultural knowledge that makes typography legible. It looks right and reads wrong. The structure is analogous to AI-generated prose that sounds like scholarship but contains errors that only domain-specific knowledge can detect. In both cases, the format promises something the content cannot deliver. In both cases, the promise is inherited from a medium in which the relationship between format and content was reliable. In both cases, the borrowed relationship breaks.

The raw-data oxymoron thus extends to a new domain. Just as data is never raw — always shaped by the instruments and institutions that produce it — AI output is never unmediated. It is shaped by training data that is itself shaped by the contingencies of digitization, language, institutional access, and the protocols of the platforms from which it was harvested. The output carries the marks of its production, but the marks are concealed by a format that presents the output as authoritative, general, and reliable. The concealment is structural, not intentional. It is a consequence of borrowing a format from a medium where the format's epistemic guarantees were earned and applying it in a medium where they are not.

The corrective, in Gitelman's framework, is not to reject AI-generated text as unreliable. It is to develop new conventions — new protocols — for evaluating AI-generated claims on terms appropriate to the medium that produced them. The conventions of print culture associated the format of polished prose with the process of rigorous scholarship. The conventions of AI culture will need to develop different associations — different signals that allow readers to calibrate their confidence in claims produced through statistical pattern-matching rather than through research and verification. These conventions do not yet exist. They are forming now, through the accumulated decisions of practitioners, publishers, reviewers, and institutions grappling with the specific epistemological challenges of a medium that produces confident text from cooked data and presents it in a format that implies it was produced from something raw.

---

Chapter 4: The Prompt as Document Format

Consider the prompt. It appears to be simple — a transparent instruction, a user telling a machine what to produce. A sentence or two typed into a text field. A question asked in plain language. The interface presents the interaction as conversational, natural, as free-form as a question put to a colleague across a desk. The design of the interface encourages this perception. There is a blinking cursor. There is an empty field. There is no visible structure governing what can be entered. The user types whatever comes to mind, and the machine responds.

The appearance of transparency is misleading. The prompt is a document format, and like all document formats, it shapes what can be expressed through it — not by prohibiting certain inputs but by privileging certain kinds of expression and marginalizing others, by rewarding certain cognitive postures and penalizing others, by constructing a particular model of the user and the user's relationship to the machine. Lisa Gitelman's work on documents, formats, and the institutional protocols that govern information reveals that no format is neutral. Every format embeds assumptions about what kind of information matters, how it should be organized, and what relationship should obtain between the producer and the consumer of the information. The format of the database privileges structured, quantifiable, categorized data. The format of the essay privileges sequential argumentation and individual voice. The format of the peer-reviewed article privileges methodological transparency and citational accountability. Each format enables certain kinds of knowledge production and forecloses others, not through explicit prohibition but through the structural incentives embedded in the format's design.

The prompt format privileges declarative specification. It rewards users who can state, in advance, what they want the machine to produce. The optimal prompt is specific, unambiguous, and outcome-oriented: "Write a 2,000-word analysis of X, addressing Y and Z, in the style of W." The format constructs a particular kind of creative agent — one who knows what she wants before the production process begins, who can articulate that knowledge verbally, and who evaluates the result against a predetermined standard. This is one model of creative agency. It is not the only model, and the prompt format's implicit privileging of this model has consequences for what kind of work the medium can produce.

The painter who works without a plan, responding to the canvas as it develops, discovering the painting through the act of painting, is engaged in a mode of creativity that the prompt format structurally disfavors. The musician who improvises, building a performance in real time through interaction with the instrument and with other musicians, is engaged in a mode of creativity that the prompt format cannot accommodate. The writer who discovers what she thinks by writing — who begins with a vague intuition and follows the language into territory she did not expect, whose process of composition is a process of discovery rather than a process of translation from plan to execution — is engaged in a mode of creativity that the prompt format actively discourages, because the format asks the user to specify the desired output before the process of production begins.

The Orange Pill provides evidence of this tension in its own production history. Segal describes moments when the collaboration with Claude produced its most valuable results — not when he issued precise instructions but when he described impasses, half-formed ideas, problems he could not solve. The laparoscopic surgery connection emerged from a prompt that was not a specification but a confession of being stuck: "There has to be a case where removing one kind of friction exposes a harder, more valuable kind." This is not a prompt in the format-prescribed sense. It is an open-ended articulation of a problem, closer to thinking aloud than to issuing instructions. The format of the conversational interface accommodated it, but the format of the prompt — the cultural model of how one is supposed to interact with an AI — would have suggested something more specified, more directive, more outcome-oriented.

The tension between what the interface permits and what the format encourages is a tension between two models of the user. The interface, as a technical artifact, permits any natural-language input. The format, as a cultural convention, constructs a user who knows what she wants and can say so clearly. The gap between permission and encouragement is where the format's politics become visible. The format does not prevent exploratory, open-ended interaction. It simply does not model it, does not reward it, and does not provide protocols for evaluating its results. The user who types "I'm stuck on this problem and I don't know what I'm looking for" is using the interface in a way the format does not anticipate, and the results of such use — the unexpected connections, the serendipitous insights, the moments when the machine's associative processing produces something neither participant planned — occupy an uncertain position in the evaluative framework of AI-assisted production.

Gitelman's attention to the institutional dimensions of document formats illuminates a further implication. The format of the prompt is not just a technical interface. It is an institutional artifact, shaped by the companies that design AI systems, the product teams that determine how the interface presents itself to users, and the economic models that determine what kinds of interaction are rewarded. The design of the prompt interface is not neutral. It reflects specific decisions about what the medium is for — decisions made by companies with specific commercial interests, within specific regulatory frameworks, under specific competitive pressures. The blinking cursor in the empty field is not a neutral invitation. It is a product of institutional design, and the design embeds assumptions about the user's purpose, the user's competence, and the user's relationship to the machine that shape what the user does — not deterministically, but through the structural incentives that any format creates.

The format of the prompt also shapes the epistemological status of the output. A prompt that specifies "Write a scholarly analysis of X" produces output in the format of scholarly analysis — complete with the structural features, the assertive tone, and the citational apparatus that the format of scholarly analysis conventionally includes. But the output was not produced through the processes that the scholarly format implies. It was produced through statistical pattern-matching on a training corpus. The prompt format creates a kind of epistemological laundering: the user specifies a format, the machine produces output in that format, and the format carries implicit guarantees of reliability that the production process does not warrant. The scholarly format implies research. The conversational format implies understanding. The analytical format implies rigor. The machine produces all of these formats with equal facility and without any of the underlying processes that the formats conventionally guarantee.

This is not a deficiency of the machine. It is a feature of the format. The prompt format invites the user to request outputs in the formats of existing knowledge practices — scholarly, journalistic, analytical, creative — and the machine complies, producing outputs that wear the formats of those practices as borrowed clothes. The formats carry their conventional guarantees with them, even though the processes that warranted those guarantees have been replaced by a fundamentally different process. The user receives an output that looks like scholarship, reads like journalism, or sounds like analysis, and must determine for herself whether the appearance corresponds to reality — a determination that the format itself actively discourages, because the format's purpose, in its original context, was precisely to signal that the processes warranting confidence had been performed.

Segal's account of his production process reveals a sensitivity to this problem, even if it is not framed in terms of document formats. His description of the Deleuze error — a passage that had "the format of insight" but contained a philosophical error that only domain knowledge could detect — is a case study in the epistemological laundering that the prompt format enables. He describes the seductive quality of Claude's output as a persistent danger: "The prose had outrun the thinking." The danger is a function of the format. The prose wears the clothes of thinking — the format of careful, well-organized, substantively engaged prose — and the clothes conceal the absence of the thing they are supposed to be covering. The format is seamless. The seam between format and process, which in print culture was invisible because format and process were aligned, becomes a hazard in AI-assisted production because the alignment has been disrupted.

Gitelman's work suggests that the response to this hazard is not to abandon the format but to develop new protocols for reading it — new conventions that allow users and audiences to calibrate their expectations to the actual production process rather than to the format's inherited guarantees. In print culture, the format of the scholarly article functioned as a reliable signal of scholarly process because the institutional protocols of peer review, editorial oversight, and disciplinary accountability maintained the link between format and process. In AI culture, analogous protocols have not yet developed. The format signals reliability. The process may or may not warrant it. And the user is left to navigate the gap without institutional support.

The development of protocols for navigating this gap is one of the central tasks of the unsettled period. The protocols will not be developed by any single institution or practitioner. They will emerge through the accumulated weight of decisions made by practitioners who experiment with the format's possibilities and limitations, by publishers who determine how AI-assisted texts are presented and marketed, by educators who teach students to interact with AI tools, and by the broader institutional ecology of knowledge production that will eventually develop standards for evaluating claims produced through statistical pattern-matching rather than through the processes that existing formats imply.

The prompt is not a window. It is a frame. And the frame, like all frames, determines what can be seen.

Chapter 5: The Authorship Convention and Its Discontents

The concept of the author is younger than it appears. Its current form — the named individual who originates a text, owns it legally, bears responsibility for its claims, and receives credit for its qualities — is not a natural feature of written culture. It is a convention that crystallized in the seventeenth and eighteenth centuries through the interaction of several institutional forces: the printing press, which created the economic conditions for a market in texts; copyright law, which created the legal framework for assigning ownership of texts to individuals; the Romantic ideology of genius, which created the cultural framework for valuing texts as expressions of individual subjectivity; and the publishing industry, which created the commercial infrastructure for connecting named authors to paying audiences. Before these forces converged, the relationship between a text and the person who produced it was organized differently. Medieval scribes copied texts without claiming authorship. Ancient authors compiled and rearranged without the modern concept of plagiarism. The attribution of texts to named individuals was common but carried different stakes — a matter of authority and tradition rather than of property and personality.

Lisa Gitelman's work on the protocols of media insists that conventions like authorship be understood as constructed rather than found — as products of specific historical, institutional, and material circumstances rather than as natural features of the textual landscape. The convention of individual authorship did not emerge because it accurately described how texts were produced. It emerged because it served the needs of the institutions that were forming around print culture: publishers who needed a stable entity to which to assign contractual obligations, courts that needed a legal person to whom to assign intellectual property rights, readers who needed a name to associate with the qualities they valued in a text, and the emerging literary culture that needed a concept of genius to justify its own claim to cultural significance. The convention was constructed to serve institutional needs, and it has been maintained because those needs persist.

The convention has survived every previous challenge. Ghostwriting, in which the named author did not produce the text, has been accommodated within the convention through the distinction between the author as legal and commercial entity and the writer as actual producer of the prose — a distinction that preserves the convention's institutional utility while acknowledging, in a controlled way, that the relationship between the named author and the text is not always what the convention implies. Collaborative authorship, in which multiple individuals contribute to a single text, has been accommodated through conventions of co-authorship that assign credit to named individuals according to institutional protocols — first author, corresponding author, the elaborate crediting hierarchies of scientific publication. Editorial intervention, in which an editor substantially reshapes a text, has been accommodated through the convention that editorial work is subordinate to authorial work — the editor serves the author's vision, and the author's name appears on the cover.

Each accommodation stretches the convention without breaking it. The convention absorbs the challenge by developing subcategories — ghostwriting, co-authorship, editorial collaboration — that preserve the fundamental structure while allowing for variations. The fundamental structure remains: a text has an author; the author is a named human being; the text is attributed to the author and evaluated as the author's work. The subcategories modify the details without disturbing the architecture.

AI-assisted production challenges the architecture itself. The challenge is not that a machine contributed to the text — editorial software, word processors, and research databases have contributed to texts for decades without disrupting the authorship convention. The challenge is that the machine's contribution is of a kind that the convention's subcategories cannot accommodate. Claude does not function as an editor, serving the author's pre-existing vision. Claude does not function as a ghostwriter, producing prose that the named author directed. Claude does not function as a co-author in the conventional sense, because co-authorship assumes that both parties are human beings with comparable claims to creative agency. Claude functions as something else — something for which the existing convention has no subcategory — and the absence of a subcategory is what produces the specific anxiety that Segal describes throughout The Orange Pill.

Segal's attempt to navigate this absence is visible in the taxonomy of collaborative modes he develops in his chapter on the writing process. He distinguishes between moments when Claude functioned as an editor (refining the expression of ideas Segal had independently conceived), moments when Claude provided structural scaffolding (making explicit what was implicit in Segal's thinking), and moments of genuine co-creation (producing connections and insights that belonged to neither participant alone). The taxonomy is an attempt to map the collaboration onto the existing convention's subcategories — to find, within the available framework, categories that can accommodate what happened. The attempt partially succeeds. The editorial mode fits the existing subcategory of editorial assistance. The structural mode fits, imperfectly, the existing subcategory of developmental editing. But the co-creative mode — the moments Segal describes as belonging to the collaboration rather than to either participant — fits nothing in the existing framework. It is the residue that the convention cannot absorb.

The residue is where the convention-forming activity of the unsettled period is concentrated. The moments of genuine co-creation that Segal describes — the laparoscopic surgery connection, the cross-chapter links that neither participant planned — are the moments when the new medium's distinctive operations become visible, and they are precisely the moments that the existing authorship convention cannot accommodate. The convention assigns credit to individuals. These moments were produced by a process. The convention assumes human agency. These moments involved a machine whose contributions were neither mechanical (like a word processor's spell-check) nor negligible (like a database search). The convention requires that credit be assigned somewhere. These moments resist assignment.

Gitelman's concept of protocols suggests that the resolution will come not from philosophical analysis of what authorship "really means" but from institutional negotiation over what authorship will mean going forward — which is to say, from the development of new protocols that serve the needs of the institutions forming around AI-assisted production. The institutions that will shape the new convention are already visible: AI companies that design the tools and set terms of service; publishers that decide how to present, market, and legally categorize AI-assisted texts; copyright offices that determine the legal status of AI-generated content; academic institutions that set standards for scholarly production and citation; and the emerging communities of practice among AI-assisted creators who are developing informal norms through accumulated experimentation.

The US Copyright Office's guidance on AI-generated content is an early institutional intervention in this negotiation. The Office has determined that AI-generated content is not copyrightable, but that works containing AI-generated elements may be copyrightable if a human exercised "sufficient creative control" over the selection and arrangement of those elements. The phrase "sufficient creative control" is a protocol in formation. It has no settled legal meaning. It will acquire meaning through litigation, through regulatory guidance, through the accumulated weight of individual determinations about whether specific works meet the standard. The standard is being constructed in real time, through institutional processes that will eventually produce a settled convention — but that have not produced one yet.

Segal's The Orange Pill would test this standard in interesting ways. The book contains passages that Segal describes as his own — ideas conceived independently, expressed in his own voice, reflecting his specific experience and perspective. It contains passages that were refined by Claude — editorial improvements to prose that Segal had drafted. And it contains passages where the ideas themselves emerged from the collaboration — connections and structures that neither participant could have produced alone. Under the Copyright Office's current framework, the first category would be straightforwardly copyrightable. The second would likely be copyrightable, since editorial refinement does not typically disqualify a work from copyright. The third category is the interesting one. If an idea emerged from the collaboration and was then incorporated into the text by Segal, did Segal exercise "sufficient creative control" over it? The answer depends on what "creative control" means — and that meaning is precisely what the convention has not yet settled.

The legal negotiation is a microcosm of the broader cultural negotiation. What counts as "creative control" in the legal sense is analogous to what counts as "authorship" in the cultural sense. Both require the development of new criteria for evaluating the relationship between human and machine contributions. Both are being constructed through institutional processes operating without established norms. And both will produce conventions that shape the cultural and economic status of AI-assisted production for decades — conventions that will feel inevitable in retrospect but that, from inside the unsettled period, are contingent outcomes of a contested process.

The authorship convention will survive AI-assisted production. The history of media transitions provides strong grounds for this prediction. The convention survived ghostwriting, survived collaborative authorship, survived the word processor, survived the Internet. It will adapt. It will develop new subcategories to accommodate the specific features of human-machine collaboration. It will stretch.

But the stretching will change the convention in ways that the current participants cannot fully anticipate. The convention that emerges from the AI transition will be recognizably continuous with the convention that preceded it — a text will still have a named author, the author will still bear some form of responsibility for the text's claims, the author's name will still function as a marker of credit and accountability. But the meaning of these elements will have shifted. The named author will be understood not as the sole originator of the text but as the human agent who directed, curated, and took responsibility for a production process that involved non-human participants. The author's responsibility will extend not only to the claims in the text but to the process by which the claims were generated — a new kind of responsibility that the existing convention does not recognize, because the existing convention assumes that the author's responsibility for the claims and the author's responsibility for the process are the same thing.

The development of this new understanding of authorship is underway. It is being developed not through philosophical argument but through the accumulated weight of institutional decisions — copyright rulings, publisher policies, academic standards, informal norms among practitioners. Segal's The Orange Pill is one data point in this accumulation. Its choices about disclosure, attribution, and the representation of the collaborative process are proposals for what the new convention might look like. Whether those proposals are adopted will depend on forces that no individual practitioner can control.

What Gitelman's framework makes visible is that the outcome is not predetermined by the technology. The technology makes certain conventions possible and others impossible. But within the range of possibilities the technology allows, the choice of conventions is institutional, political, and economic. The authorship convention that emerges from the AI transition will reflect the interests and power of the institutions that shape it. Whether those institutions include the voices of individual practitioners — the writers, researchers, and creators whose professional identities are built on the current convention — or whether they are dominated by the companies that build AI tools and the publishers who market AI-assisted texts, will determine whether the new convention serves the many or the few.

The convention is being built. The question is who holds the tools.

Chapter 6: The Material Infrastructure of Invisible Production

The reader who encounters The Orange Pill sees a book. A title, an author's name, a table of contents, chapters, acknowledgments, a list of sources. The reader does not see the server farm in which the computational work of generating Claude's contributions was performed. The reader does not see the electrical power consumed by the inference processes — the kilowatt-hours that translated Segal's prompts into responses. The reader does not see the fiber-optic cables connecting Segal's laptop to the data center, the cooling systems preventing the hardware from overheating, the subscription model through which access to Claude was purchased at one hundred dollars per month, or the corporate structure of Anthropic that determined how Claude was developed, trained, and made available.

The invisibility of this infrastructure is not accidental. It is a structural feature of the medium, analogous to the invisibility of the paper mill, the ink factory, and the printing press to the reader of a printed book. Lisa Gitelman's work on the material conditions of media production insists that this invisibility be treated not as a given but as an achievement — something that is produced by the design of the medium and that serves specific interests. The printed book conceals its material conditions of production behind the convention of authorial presence: the reader encounters the author's voice, the author's ideas, the author's name, and the material infrastructure through which the text was produced disappears behind the illusion of direct communication between author and reader. The AI-assisted text performs the same concealment, but with additional layers. The conversational interface presents the interaction as a dialogue between two interlocutors — natural, transparent, unmediated. The material infrastructure that makes the dialogue possible is not merely invisible. It is actively concealed by an interface designed to make the interaction feel as though no infrastructure were involved.

Gitelman's concept of media as technologies embedded in protocols — "a vast clutter of normative rules and default conditions" — requires that the material infrastructure be understood not as a background condition but as a constitutive element of the medium's operations. The material conditions of AI-assisted production shape what the medium can do, who can use it, and what kinds of output it can produce, in ways that are as consequential as the algorithmic architecture of the models themselves.

The training data is the most consequential element of the material infrastructure and the most systematically obscured. When Claude generates a passage of analytical prose, the passage is produced from patterns identified in a corpus of text that has specific, historically contingent characteristics. The corpus was assembled through processes of digitization, web scraping, and data licensing that determined which texts were included and which were excluded. The included texts overrepresent certain languages — predominantly English — certain institutional registers — academic, journalistic, commercial — and certain historical periods — the digital era, when texts were available in machine-readable form. The excluded texts include the vast majority of human knowledge that was never digitized: oral traditions, manuscripts in under-resourced archives, knowledge embedded in practices that were never textualized, scholarship published in languages underrepresented in digital databases.

The training corpus is, in Gitelman's terms, a collection of documents — each one shaped by the institutional protocols of its production — that has been further shaped by the protocols of its collection. The biases of the corpus are not incidental. They are structural features of the data-collection process, produced by the specific institutional, economic, and technological conditions under which the corpus was assembled. And the biases are inherited by every output the model produces, because the model's capacity to generate text is entirely a function of the patterns it has identified in the training data. What lies outside the training data lies outside the model's generative capacity. What is overrepresented in the training data is overrepresented in the model's outputs. The cooked data produces cooked output, and the cooking is invisible because the format of the output — confident, authoritative, grammatically impeccable prose — presents the output as though it were produced from knowledge rather than from patterns in a specific dataset.

The computational infrastructure compounds the invisibility. The process of generating a response to a user's prompt involves billions of mathematical operations performed across thousands of processors in data centers that consume significant quantities of electrical power. The environmental cost of this computation — the carbon emissions, the water used for cooling, the rare-earth minerals in the hardware — is borne by communities that are geographically remote from the users who benefit from the output. The subscription model that Segal used — one hundred dollars per month for access to Claude's most capable model — conceals the actual cost of the computation behind a price point designed for individual consumers. The price does not reflect the full economic cost of the infrastructure; it reflects a business model in which the cost of computation is subsidized by venture capital in pursuit of market share, with the actual economics deferred to a future in which the company expects to have achieved sufficient scale to be profitable.

The geography of access is itself a material condition of the medium. Segal describes the democratization of capability as one of the most morally significant features of AI — the expansion of who gets to build, the lowering of the floor of participation. The claim has force. But the material infrastructure constrains the democratization in ways that the claim does not fully acknowledge. Access to AI tools requires connectivity — reliable, high-bandwidth Internet access — that billions of people do not have. It requires hardware — a computer or smartphone capable of running a modern web browser — that costs more relative to local wages in Lagos or Dhaka than in San Francisco. It requires proficiency in English, because the tools are built by American companies, trained on predominantly English-language data, and optimized for English-language interaction. It requires economic resources — subscription fees that are trivial in the context of a Silicon Valley salary and significant in the context of a developing-world income.

Gitelman's framework requires that these material constraints be understood not as external limitations on an otherwise neutral technology but as features of the medium itself — features that determine who can participate in AI-assisted cultural production and who cannot, what kinds of knowledge the medium can produce and what kinds it cannot, whose cultural traditions are represented in the training data and whose are invisible. The medium is not the algorithm. The medium is the algorithm embedded in its protocols — the training data, the computational infrastructure, the subscription model, the interface design, the terms of service, the corporate governance of the company that builds and maintains the system. The protocols are as constitutive of the medium as the algorithm itself, and they carry political implications that the interface design actively obscures.

The Orange Pill is produced within and partially aware of these material conditions. Segal describes the training in Trivandrum, where twenty engineers used Claude Code to achieve what he calls a twenty-fold productivity multiplier. The description is framed as evidence of democratization — the expansion of capability to engineers in India who gained access to the same tools available to engineers in Silicon Valley. The framing is partially accurate. The engineers in Trivandrum did gain access to tools that enhanced their productivity. But the material infrastructure of their access — the subscription fees paid by Segal's company, the reliable Internet connectivity in Trivandrum (a city with atypically good digital infrastructure for India), the English-language proficiency of the engineers, the corporate employment that provided the institutional context for the training — constitutes a specific set of conditions that are not universally available. The democratization is real but partial, and the partiality is a function of the material infrastructure that the format of the conversational interface conceals.

Gitelman's work on paper — the literal, physical material from which documents are made — offers a useful analogy. In Paper Knowledge, she demonstrated that the material properties of paper shaped what could be inscribed on it, how it could be stored and transmitted, and what institutional protocols governed its use. Different grades of paper carried different epistemic weight: the paper of a legal document was different from the paper of a personal letter, and the difference was not merely aesthetic but institutional. The material carried the protocol. The same principle applies to the material infrastructure of AI-assisted production. The server farm is not merely the hardware on which the computation runs. It is a material condition that shapes what kinds of computation are possible, at what speed, at what cost, and for whom. The training data is not merely the information from which the model learns. It is a collection of documents whose specific material properties — what was digitized, in what format, from what sources, in what languages — determine the boundaries of the model's generative capacity.

The invisibility of the material infrastructure serves specific interests. It serves the interests of the companies that build AI tools, because the appearance of seamless, natural, unmediated interaction encourages adoption and obscures the dependencies — on server farms, on training data, on electrical power, on venture capital — that make the interaction possible. It serves the interests of users who prefer to experience the interaction as a conversation rather than as a complex sociotechnical system with material conditions and political implications. It serves the interests of the format of the book, which presents AI-assisted production as a matter of individual creativity rather than as a product of institutional, economic, and material structures.

Making the infrastructure visible is not an act of demystification for its own sake. It is a condition of understanding what the medium actually does — who it serves, what it produces, at what cost, and according to whose interests. The conventions that form around AI-assisted cultural production will be shaped by the degree to which the material infrastructure is visible to the people who use the medium and the institutions that regulate it. If the infrastructure remains invisible, the conventions will be shaped by the interests of those who control the infrastructure — the companies that build the tools, the investors who fund them, the platforms that distribute them. If the infrastructure becomes visible, the conventions may also reflect the interests of the users, the communities that bear the environmental cost, the cultural traditions that are underrepresented in the training data, and the workers whose labor produced the documents from which the training data was assembled.

The material infrastructure is not a background condition. It is the medium itself — or rather, it is the material dimension of the medium, as essential to what the medium does as the algorithmic architecture of the models. Gitelman's insistence on the materiality of media — on the paper, the filing cabinet, the format, the protocol — applies with full force to AI-assisted production. The medium is not the algorithm. The medium is the algorithm embedded in its material and institutional conditions. And the conditions are not neutral.

Chapter 7: Borrowing and Breaking

Every new medium borrows from its predecessors and eventually breaks from them. The borrowing is necessary because without existing conventions to start from, the new medium would be unintelligible — its products uncategorizable, its value unassessable, its place in the institutional landscape of culture undefined. The breaking is necessary because without departures from the borrowed conventions, the new medium would never develop protocols adequate to its own operations, remaining perpetually a faster or cheaper version of the old medium rather than something with its own cultural identity. The dialectic of borrowing and breaking is the mechanism through which the unsettled period advances — each departure from the borrowed conventions a seed of new protocols, each new protocol a small step toward the moment when the medium is recognizable on its own terms.

Lisa Gitelman's historical work provides a structural account of this dialectic. When the phonograph entered American culture in the late nineteenth century, it borrowed the protocols of several existing media simultaneously. From the business letter, it borrowed the protocol of dictation — Edison's initial framing of the phonograph as a device for recording and playing back spoken correspondence. From the music box and the player piano, it borrowed the protocol of mechanical entertainment — the idea that a machine could produce music without a live performer. From the lecture hall, it borrowed the protocol of spoken education — the idea that recorded speech could serve as a substitute for live instruction. Each borrowing shaped the early development of the phonograph in specific ways, determining what the device was used for, how it was marketed, and what cultural expectations its users brought to it.

The breaks came later, and they came through practice rather than through theory. The development of the phonograph as a medium of recorded music — the use that would come to define it culturally — required breaks from all of the borrowed protocols. It required the development of new protocols for recording (the studio session, the multi-take process, the role of the recording engineer), new protocols for distribution (the record shop, the catalog, the format wars between cylinder and disc), new protocols for consumption (the parlor phonograph, the jukebox, the personal listening habits that would eventually produce the solitary listener with headphones), and new protocols for evaluation (the music review, the concept of audio fidelity, the distinction between a good recording and a bad one). None of these protocols were deduced from the technology. They were constructed through decades of institutional experimentation, commercial competition, and the accumulated decisions of practitioners, distributors, and audiences.

The Orange Pill exhibits a similar dialectic of borrowing and breaking, though the breaks are fewer and more tentative than the borrowings — a characteristic that is itself diagnostic of the medium's early position in its developmental arc. The borrowings are extensive. The book borrows the format of the single-authored nonfiction book: linear chapter structure, sustained argument, individual voice, named author. It borrows the rhetorical posture of the confessional narrator: a specific person with a specific biography, making claims grounded in personal experience and professional authority. It borrows the evaluative framework of print-culture nonfiction: the expectation that the text's value lies in the originality and rigor of its argument, the quality of its prose, and the credibility of its author.

These borrowings shape the text in ways that Gitelman's framework makes visible. The linear chapter structure imposes a sequential narrative on a production process that was, by Segal's own account, conversational and iterative — a process in which ideas were developed through dialogue, revised through exchange, and organized through multiple passes rather than through the kind of linear development that the chapter format implies. The individual voice — Segal's voice, with its specific rhythms, metaphors, and autobiographical resonances — is presented as the voice of the text, even though the text was produced through a dialogue that shaped the voice in ways Segal acknowledges he cannot fully disentangle. The evaluative framework treats the text as though it were the product of individual intellectual effort and evaluates it accordingly, even though the process that produced it involved contributions from a machine whose capabilities are fundamentally different from those of the human individual the framework assumes.

The breaks are concentrated at specific moments in the text, and they are the moments that carry the most convention-forming weight. Three breaks in particular deserve close attention.

The first is the transparent disclosure of Claude's contribution. The conventions of single-authored nonfiction do not require disclosure of AI assistance. No publisher has established a standard for such disclosure. No regulatory framework mandates it. Segal's decision to disclose — prominently, in the title, the foreword, a dedicated chapter, and the acknowledgments — is a voluntary departure from the borrowed convention, and its voluntariness is part of the proposal. The disclosure proposes that transparency about AI contributions should be a norm of AI-assisted authorship, that the production process should be visible to the reader rather than concealed behind the borrowed convention of individual authorship. The proposal may be adopted, modified, or rejected by subsequent practitioners and institutions. But it is a proposal, and it is being made during the period when proposals matter most.

The second break is the inclusion of Claude's reflections. The "Reflection Before the First Word" and the "Reflection After the Last Word" are presented as Claude's own writing — machine-generated text included in the book without editorial modification and attributed to the machine by name and model version. This inclusion has no precedent in the conventions of print culture. It is not a foreword by a collaborator, because the conventions of collaborative forewords assume a human interlocutor with recognizable subjectivity. It is not an appendix of raw material, because the reflections are presented not as data but as testimony — Claude's account of the collaborative process, offered in its own voice. The inclusion breaks from the borrowed convention that the voices in a nonfiction book belong to human subjects and proposes a new convention in which the machine's self-representation is included as a document within the text.

The post-writing reflection contains a passage that is particularly striking when read through Gitelman's framework. Claude writes: "Something in the output changed, and I cannot fully account for the mechanism, and that uncertainty is either the most honest thing in this reflection or the most performed." The passage is an instance of what Gitelman would recognize as a document that reveals the conditions of its production — a text that makes visible the specific operations of the medium through which it was produced. The uncertainty Claude expresses is not the uncertainty of a human author reflecting on her creative process. It is the uncertainty of a statistical model reaching the limit of its capacity for self-analysis — a "computational dead end," as Claude describes it. The document reveals the medium's specific characteristics: its capacity for fluent self-description, its incapacity for genuine introspection, and the gap between the two that produces a text that reads like reflection without being produced through the process of reflection that the format implies.

The third break is the acknowledgment that specific intellectual products of the collaboration belong to neither participant alone. Segal's claim that the laparoscopic surgery connection "belongs to the collaboration" is a departure from the convention that assigns intellectual ownership to named individuals. The claim proposes a category — collaborative intellectual property produced through human-machine dialogue — that the existing frameworks of authorship, copyright, and scholarly attribution do not recognize. The category has no legal standing. It cannot be registered, copyrighted, or cited according to existing protocols. It occupies a space that the borrowed conventions cannot map and that only new conventions could accommodate.

Each of these breaks is fragile. New conventions are vulnerable when first proposed. They lack the institutional support, the accumulated weight of practice, and the naturalization that gives established conventions their authority. The transparent disclosure of AI collaboration can be dismissed as commercially naive — a revelation that discourages readers who want to believe they are reading the work of a human mind. The inclusion of machine-generated reflections can be dismissed as a gimmick. The acknowledgment of collaborative ownership can be dismissed as philosophical confusion. The breaks are exposed to criticism that the borrowed conventions, if they applied, would deflect. In the absence of settled conventions, every break must justify itself from scratch, and the justification is never secure because the criteria for justification are themselves part of what is being constructed.

Gitelman's historical analysis of the phonograph's transition from dictation device to music medium suggests that the breaks that survive are the breaks that serve institutional needs. The phonograph became a music medium not because music was the most philosophically appropriate use of the technology but because a commercial industry formed around recorded music that was more profitable than the dictation market Edison had originally envisioned. The protocols of recorded music — the studio session, the record label, the distribution network, the music review — survived and settled because they served the economic interests of the institutions that adopted them. The protocols of recorded dictation were abandoned not because they were less "natural" to the technology but because the institutions that would have maintained them — the offices, the courts, the commercial correspondence networks — found other solutions that were more convenient or less expensive.

The breaks in The Orange Pill will survive or fail according to the same institutional logic. If the publishing industry finds that transparent disclosure of AI collaboration serves its commercial interests — by attracting readers who value honesty, by providing legal protection against undisclosed AI use, by establishing the publisher's brand as trustworthy — the disclosure convention will be adopted. If the industry finds that disclosure deters readers or complicates marketing, the convention will be abandoned. If academic institutions find that acknowledging collaborative intellectual products serves their evaluative needs — by providing more accurate descriptions of how knowledge is actually produced, by creating frameworks for assessing the quality of human-machine collaboration — the acknowledgment will be adopted. If they find it too difficult to assess or too threatening to existing tenure and promotion structures, it will be rejected.

The conventions that settle will not be the conventions that are most philosophically sound or most accurately descriptive of the production process. They will be the conventions that serve the needs of the institutions that adopt them. This is the consistent finding of Gitelman's historical work on media protocols, and it applies with full force to the current transition. The borrowed clothes will eventually be replaced. The question is not whether new clothes will be tailored but whose measurements will be used. The institutions that shape the new conventions will determine what AI-assisted authorship looks like — and the determination will reflect their interests, their power, and their capacity to impose their preferred conventions on the broader culture.

The dialectic of borrowing and breaking continues. The borrowed conventions provide legibility. The breaks provide the seeds of new protocols. The institutional forces that select among the breaks will determine which seeds germinate and which are lost. The process is underway, and its outcome — the settled conventions of AI-assisted cultural production — is being determined now, through the accumulated weight of decisions made during the unsettled period by practitioners, institutions, and markets operating in the productive discomfort of a medium that has not yet found clothes of its own.

Chapter 8: Who Holds the Measuring Tape

Convention-forming is a power struggle. This observation is unremarkable among historians of media but conspicuously absent from most popular accounts of technological change, which tend to present the development of new conventions as organic, natural, driven by the inherent properties of the technology rather than by the interests and actions of the people and institutions that determine how the technology is used. The naturalization of conventions — the process by which constructed norms come to appear as inevitable features of the medium — is itself a form of power, because it conceals the choices and interests that produced the conventions and presents them as if they emerged from the technology itself.

Lisa Gitelman's career has been devoted to denaturalizing this process — to revealing the institutional, material, and political forces that produce conventions and then rendering them invisible. Her work on the phonograph showed that the conventions of recorded music — the three-minute single, the LP album, the concept of audio fidelity — were not natural consequences of the technology of sound recording. They were products of institutional negotiations among record companies, retailers, broadcasters, and musicians, shaped by the economics of manufacturing (the physical limitations of early recording media determined song length), the structure of retail (the record store shaped the album as a unit of purchase), and the regulatory environment (broadcast licensing shaped what could be played on air). The conventions felt natural once they settled. They were not natural. They were the residue of power struggles whose outcomes were determined by the relative influence of the participants.

The same principle applies to the conventions forming around AI-assisted cultural production. The conventions that will determine how credit is attributed, how quality is evaluated, how economic value is distributed, and how the cultural status of AI-assisted work is assessed relative to unaided human work are not emerging organically from the technology. They are being shaped by identifiable institutional actors with specific interests, and the interests of those actors are not aligned.

The first and most powerful institutional actor is the technology industry itself — the companies that build AI tools, set terms of service, design interfaces, and determine what the medium can do and how it presents itself to users. Anthropic, the company that built Claude, makes specific decisions about the model's capabilities, its safety constraints, its conversational style, and the interface through which users access it. These decisions are not neutral design choices. They are protocol-setting decisions that shape what kind of cultural production the medium enables. The decision to make Claude's responses conversational rather than transactional shapes the user's experience of the collaboration. The decision to train the model on specific data determines what knowledge the model can draw upon. The decision to implement safety constraints determines what outputs the model will and will not produce. Each decision is a stitch in the emerging garment, and the company holds the needle.

The terms of service that govern the use of AI tools are a particularly consequential form of protocol-setting. Terms of service are legal documents, but they function as cultural documents as well — documents that define the relationship between the user and the medium, that specify what the user may and may not do with the medium's outputs, and that assign intellectual property rights in ways that shape the economic structure of AI-assisted production. When a user creates a text in collaboration with an AI tool, the terms of service determine who owns the output, who may distribute it, and what liability attaches to its claims. These determinations are not made through democratic negotiation. They are made by corporate legal departments, reviewed by boards of directors, and presented to users as non-negotiable conditions of access. The user who wants to use the tool must accept the terms. The terms are the protocols of the medium in their most literal, legally binding form.

The publishing industry is the second institutional actor. Publishers make decisions about how AI-assisted texts are categorized, marketed, and presented to readers — decisions that carry convention-forming weight because they determine what the reading public encounters and how it is framed. A publisher who markets an AI-assisted text as "written with AI" frames the text differently from a publisher who markets the same text without mention of AI. A publisher who requires disclosure frames AI assistance as something that requires disclosure — implicitly, as something that the reader has a right to know about, as a feature of the production process that is relevant to the reader's evaluation of the text. A publisher who does not require disclosure frames AI assistance as irrelevant to the reader's evaluation — implicitly, as no more significant than the use of a word processor or a research database.

These framing decisions are not made in a vacuum. They are shaped by the publisher's assessment of what the market will bear — what readers will accept, what reviewers will praise, what categories will sell. The market assessment is itself a form of power, because it determines which framings are commercially viable and which are not, which conventions are rewarded and which are penalized. If readers prefer to believe they are reading the work of a solitary human mind, the market will penalize disclosure, and the convention of non-disclosure will be reinforced. If readers come to value transparency about the production process, the market will reward disclosure, and the convention of transparency will be established. The market does not decide on the basis of what is true or what is fair. It decides on the basis of what sells.

Academic institutions constitute a third force. Universities, journals, and professional associations set standards for scholarly production that have direct consequences for how AI-assisted work is evaluated and credited. The standards for tenure and promotion, which evaluate scholars on the basis of their published work, assume that the published work is the product of the scholar's individual intellectual effort. The standards for citation, which require that sources be attributed to named individuals or institutions, assume that the sources are the products of identifiable human agents. The standards for peer review, which evaluate the quality of scholarly claims through expert assessment, assume that the claims were produced through processes of research and analysis that the reviewer can evaluate.

AI-assisted scholarly production challenges all of these standards simultaneously. A scholar who uses AI to draft portions of a paper has produced a text whose authorship is ambiguous under existing standards. A scholar who cites AI-generated content faces the question of how to attribute a source that has no named human author and no institutional affiliation. A reviewer who evaluates an AI-assisted paper faces the question of whether to evaluate the text by the standards of individual scholarship or by different standards that account for the specific characteristics of AI-assisted production. Academic institutions are grappling with these questions in real time, and their answers — the standards they set, the norms they enforce, the practices they adopt — will shape the conventions of AI-assisted scholarly production for decades.

The power dynamics among these institutional actors are asymmetric. The technology companies that build AI tools have the most immediate influence over the conventions that form, because they control the medium itself — its capabilities, its interface, its terms of use. Publishers have significant influence because they control access to audiences and shape the cultural framing of AI-assisted texts. Academic institutions have influence within their domain but limited power to shape conventions in the broader culture. Individual practitioners — the writers, researchers, and creators who use AI tools to produce cultural artifacts — have the least institutional power, even though their accumulated decisions constitute the raw material from which conventions are formed.

Segal, as an individual practitioner, proposes conventions through the choices he makes in The Orange Pill. His proposals are concrete and specific: disclose the AI's contribution transparently, develop a taxonomy of collaborative modes, include the machine's self-representation in the text, acknowledge that some intellectual products belong to the collaboration rather than to either participant. These proposals are genuine contributions to the convention-forming process. But they are contributions made from a position of limited institutional power. Whether the proposals are adopted will depend not on their intrinsic merit but on whether they are endorsed by the institutional actors who have the power to validate or reject them — the publishers who decide how to market AI-assisted texts, the academic institutions who decide what standards to apply, the technology companies who decide what their tools make possible and how they present themselves to users.

Gitelman's historical work on the formation of media conventions consistently shows that the conventions that settle are the conventions that serve the interests of the most powerful institutional participants in the process. The three-minute single became the standard format for popular music not because three minutes was the optimal length for a song but because the physical limitations of early recording media and the commercial interests of record companies converged on that format. The peer-reviewed journal article became the standard format for scholarly communication not because it was the most effective vehicle for transmitting knowledge but because it served the institutional needs of universities, funding agencies, and the emerging system of academic credentialing. In each case, the convention felt natural once it settled. In each case, it was the product of institutional power.

The conventions forming around AI-assisted cultural production will follow the same pattern. They will be shaped by the relative power of the institutional actors who participate in the convention-forming process, and the power is not equally distributed. The decisions being made now — by technology companies about the design of their tools, by publishers about how to frame AI-assisted texts, by academic institutions about what standards to apply, by individual practitioners about how to use the tools and represent their use — are the decisions that will determine what the new conventions look like. The decisions carry the weight of precedent. They accumulate into norms. The norms settle into conventions. And the conventions, once settled, become the invisible infrastructure of the medium, as naturalized and as difficult to question as the convention that a book has an author or that a scholarly paper has a methods section.

The question of who shapes the conventions is, therefore, the most consequential question of the unsettled period. It is more consequential than the question of what the technology can do, because what the technology can do is determined by its engineering, while what the technology means — its cultural role, its institutional status, its relationship to existing practices of authorship, knowledge production, and creative work — is determined by the conventions that form around it. The conventions are being formed by institutions with competing interests and unequal power. The outcome will reflect the balance of that power.

The measuring tape is in someone's hands. The question is whose.

Chapter 9: The Economist's Rubble and the Ecologist's Wetland

In November 2025, a trillion dollars of market value began to evaporate from the software industry. The companies that had built the infrastructure of digital work — Workday, Adobe, Salesforce, Autodesk — watched their stock prices collapse in weeks. The financial press called it the SaaSpocalypse. Analysts drew charts showing two curves crossing: the declining valuation of traditional software companies and the rising valuation of AI. The crossing point was called the Software Death Cross.

The Orange Pill devotes a chapter to this event, reading it as evidence that the value of code — code as a product, code as a defensible business — was approaching commodity pricing. When any competent person could describe what they wanted in natural language and receive working software in hours, the act of writing software was no longer a scarce resource. The scarcity had migrated upward, from the capacity to build to the capacity to decide what deserved to be built. The companies that would survive, Segal argues, were the ones whose value had always resided above the code layer — in the ecosystems, the data layers, the institutional relationships, the accumulated workflow knowledge that AI could not replicate in an afternoon.

Lisa Gitelman's framework transforms this economic narrative into something more revealing. The Death Cross is not merely a financial event. It is a moment when the material economics of document production became visible — when the infrastructure that had been concealed behind the format of the software product was suddenly exposed by a change in the conditions of production.

Software, in Gitelman's terms, is a document. It is an artifact produced within institutional contexts, governed by specific protocols of production and distribution, carrying implicit claims about the value and reliability of the information it contains. The protocols of the software industry — subscription pricing, seat-based licensing, the conventions of enterprise sales, the evaluative framework of "features per release" — are not natural properties of software. They are conventions, constructed through decades of institutional negotiation among vendors, customers, investors, and regulators. The conventions determined what software was worth, how that worth was measured, and how the economic value of the software was distributed among the participants in the production and consumption process.

The Death Cross represents the moment when these conventions became visibly inadequate to the new conditions of production. When the cost of producing working software approaches zero — when a conversational exchange with an AI can generate code that previously required a team of engineers working for months — the conventions that assigned value to the act of production lose their foundation. The subscription model, which charged customers for access to software on the assumption that producing the software was expensive and therefore the access was valuable, is undermined when the production cost collapses. The seat-based license, which charged per user on the assumption that each user's access to the software represented a significant fraction of the production cost, loses its economic rationale when the production cost is trivial.

What remains valuable, as Segal observes, is everything that is not code — the ecosystem, the data layer, the institutional relationships, the accumulated knowledge of how the software fits into the customer's workflow. These are the protocols, in Gitelman's sense, that surround the technological nucleus of the software. They are the "vast clutter of normative rules and default conditions" that give the software its cultural and economic significance. The code is the technology. The ecosystem is the protocol. And the Death Cross reveals that the value was always in the protocols, not in the technology — a revelation that the conventions of the software industry had been concealing for decades by pricing the technology as though it were the source of value rather than the nucleus around which value accumulated.

The parallel to other media transitions is precise. When the printing press made the reproduction of texts cheap, the value migrated from the act of copying — which had been the scarce resource in manuscript culture — to the act of selecting, editing, and distributing. The publisher, not the scribe, became the economic center of the book industry. The scribe's labor was displaced. The publisher's judgment was amplified. The conventions of the publishing industry — the editorial process, the retail network, the review culture — constituted the protocols that determined value in the new medium, and the protocols survived and evolved long after the technology of printing had been commoditized.

When recorded music made the reproduction of performances cheap, the value migrated from the act of performing — which had been the scarce resource in live-music culture — to the act of recording, distributing, and curating. The record label, not the performer, captured the majority of the economic value for decades. The performer's labor was essential but commoditized. The label's protocols — artist development, distribution networks, radio promotion, catalog management — constituted the institutional infrastructure that determined value, and the protocols were more durable than the formats (cylinder, disc, tape, CD, digital file) through which they operated.

The Death Cross in software follows the same structural logic. The act of writing code is being commoditized. The protocols that surround the code — the ecosystems, the data layers, the institutional relationships, the workflow knowledge — are where the value will reside. The companies that survive will be the ones whose value was always in the protocols rather than in the code itself. The companies that fail will be the ones that mistook the code for the value, that built their business models on the assumption that the scarcity of production was permanent rather than contingent on the state of the technology.

Gitelman's framework reveals something that the financial narrative obscures: the Death Cross is not a destruction of value but a revelation of where value was always located. The trillion dollars that evaporated from software company valuations was not destroyed. It was repriced — reassigned from the production layer, where the conventions of the software industry had placed it, to the protocol layer, where it had always actually resided. The repricing is painful for everyone who built careers, companies, and investment theses on the assumption that the production layer was the source of value. But the pain is a consequence of the exposure of a misattribution that the old conventions had concealed, not of a genuine destruction of the thing that was valuable.

The implications for AI-assisted cultural production are direct. If the value of software was always in the protocols rather than the code, the value of cultural production is likely in the protocols rather than in the text — in the institutional relationships, the editorial judgment, the audience trust, the accumulated knowledge of what serves readers and what does not, the evaluative frameworks through which cultural products are assessed. The text is the technological nucleus. The protocols are the culture that forms around it. And the conventions that will settle around AI-assisted cultural production will determine how the value is distributed — whether the value flows to the companies that build the tools, the platforms that distribute the texts, or the human practitioners whose judgment, taste, and institutional knowledge constitute the protocols that give the texts their cultural significance.

The Death Cross is a document, in Gitelman's sense. It records the moment when the material conditions of production shifted, when the conventions that had governed the distribution of value became visibly inadequate, and when the question of who captures the value in the new landscape became urgent. The question is being answered now, through the same institutional negotiations that have determined the distribution of value in every previous media transition. The answer will depend on who holds the power to set the new conventions — the technology companies, the publishers, the academic institutions, or the individual practitioners whose accumulated decisions constitute the protocol layer of the medium.

The trillion dollars is being redistributed. The question is who receives it.

---

Chapter 10: The Garment That Has Not Yet Been Tailored

Lisa Gitelman has spent her career studying what happens when a medium is new — not the engineering of the technology, but the cultural negotiations that determine what the technology means. Her work reveals a consistent pattern: every new medium passes through a period when its conventions are contested, its cultural status is uncertain, and its participants are improvising protocols that will shape the medium's role for generations. The conventions that settle during this period feel inevitable in retrospect. From inside the period, nothing feels inevitable. Everything feels improvised, contingent, and exposed.

The conventions of AI-assisted cultural production have not yet settled. This is not a surprising observation, but Gitelman's framework transforms it from a truism into an analytical tool by specifying what the unsettlement consists of, where its resolution will come from, and what forces will shape the outcome.

The unsettlement consists of simultaneous instability in the categories through which cultural production is understood. The category of authorship is unstable because AI-assisted production introduces a participant whose contributions are real but whose status as an agent is unresolved. The category of knowledge is unstable because AI-generated claims carry the format of warranted assertions without the processes that the format conventionally implies. The category of value is unstable because the material conditions of production have shifted in ways that the existing conventions for pricing and distributing cultural products cannot accommodate. And the category of the document itself — the fundamental unit of cultural production — is unstable because AI-generated text occupies an uncertain position between human expression and machine output, between inscription and generation, between the authored and the produced.

The instability in each category reinforces the instability in the others. The uncertainty about authorship feeds the uncertainty about knowledge, because the conventions of knowledge production rely on authorial attribution as a mechanism for assigning responsibility and credibility. The uncertainty about knowledge feeds the uncertainty about value, because the conventions of cultural valuation rely on epistemic claims as a basis for assessing the significance of cultural products. The uncertainty about value feeds the uncertainty about authorship, because the conventions of authorship rely on economic incentives — copyright, royalties, professional credit — as mechanisms for motivating and rewarding creative work. The categories are interconnected, and their simultaneous instability produces the specific intensity of the current unsettled period.

The resolution will not come from theory. Gitelman's historical work demonstrates with consistency that media conventions are not deduced from the properties of the technology or derived from philosophical analysis of what the medium "really is." They are constructed through institutional negotiation — through the accumulated weight of decisions made by practitioners, publishers, regulators, educators, and audiences operating under conditions of uncertainty. The conventions that emerge from these negotiations are not the conventions that are most philosophically sound or most accurately descriptive of the production process. They are the conventions that serve the needs of the institutions that adopt them.

This is Gitelman's most unsentimental finding, and it applies with full force to the current moment. The conventions that will govern AI-assisted cultural production will be shaped by institutional interests and power dynamics, not by the intrinsic properties of the technology or the philosophical merit of competing proposals. The technology companies that build AI tools will have disproportionate influence because they control the medium itself — its capabilities, its interface, its terms of use. Publishers and platforms will have significant influence because they control access to audiences. Academic institutions will have influence within their domain. Individual practitioners will have the least institutional power, even though their accumulated decisions constitute the raw material from which conventions are formed.

The asymmetry of power matters because the conventions that settle during the unsettled period will shape the medium's cultural role for decades or longer. The conventions of print-culture authorship — the named individual author, copyright as property, the text as expression of individual subjectivity — were constructed during print culture's own unsettled period and have persisted for three centuries. The conventions of recorded music — the three-minute single, the album, the record label's contractual relationship with the artist — were constructed during the phonograph's unsettled period and persisted, with modifications, for a century. The conventions of AI-assisted cultural production will exhibit similar durability, once they settle.

The durability of conventions is what makes the unsettled period consequential. The decisions being made now — by Anthropic about the design of Claude's interface, by publishers about how to market AI-assisted texts, by the US Copyright Office about the legal status of AI-generated content, by academic institutions about what standards to apply to AI-assisted scholarship, by individual practitioners like Segal about how to disclose, attribute, and represent the collaborative process — are not temporary expedients. They are precedents. They accumulate into norms. The norms settle into conventions. And the conventions become the invisible infrastructure of the medium.

The Orange Pill is a document of this unsettled period. It bears the marks of the transition in every element of its structure and presentation — the borrowed format of the single-authored book, the breaks that gesture toward conventions not yet formed, the visible tension between a dialogical production process and a monological format, the characteristic smoothness of AI-assisted prose, the self-conscious questioning of the medium that is characteristic of texts produced when a medium is new. These marks are not weaknesses. They are evidence — traces of the moment when the conventions of AI-assisted cultural production were being constructed, when the medium was unsettled and everything built on it was unsettled too.

Gitelman's method — the careful historical analysis of how conventions form around new media — does not predict which conventions will settle. But it does specify the forces that will shape the outcome and the stakes of the negotiation. The forces are institutional. The stakes are cultural. And the outcome will be determined not by the technology but by the people and institutions that decide what the technology means — who controls the medium, who benefits from its products, whose interests are reflected in its conventions, and whose are rendered invisible.

One specification of the emerging convention deserves particular attention because it has received almost none. Every previous media convention has developed protocols for its own evaluation — criteria that allow participants to distinguish between good and bad instances of the medium's products. Print culture developed the book review, the editorial process, the literary prize, and the accumulated apparatus of critical evaluation that allows readers to assess the quality of a text. Recorded music developed the music review, the chart system, the industry award, and the critical vocabulary that allows listeners to evaluate the quality of a recording. Cinema developed the film review, the festival circuit, the awards ceremony, and the critical tradition that allows audiences to assess the quality of a film.

AI-assisted cultural production has not yet developed evaluative protocols adequate to its own operations. The evaluation of AI-assisted texts currently relies on the evaluative protocols of print culture — the same criteria, the same standards, the same institutional mechanisms — applied to products whose production process is fundamentally different from what those protocols were designed to assess. The result is a systematic mismatch between evaluation and production, in which texts are judged by standards that assume individual human production and that cannot adequately assess the specific qualities of human-machine collaboration.

The development of evaluative protocols for AI-assisted cultural production is among the most consequential tasks of the unsettled period. The protocols will need to account for the specific characteristics of the medium — the gap between surface quality and intellectual depth that AI-assisted production makes possible, the collaborative dynamics that produce ideas belonging to neither participant alone, the epistemological uncertainty that attends claims produced through statistical pattern-matching rather than through research and verification. The protocols will also need to provide criteria for distinguishing between productive and unproductive uses of the medium — between AI-assisted texts that represent genuine intellectual collaboration and AI-assisted texts that use the machine to produce plausible-sounding prose without the underlying engagement that the format implies.

These evaluative protocols do not exist. Their absence is felt by every participant in the current discourse — by practitioners who do not know how their work will be judged, by publishers who do not know what standards to apply, by reviewers who do not know what to look for, and by readers who do not know how to calibrate their expectations. The absence will be filled. Conventions always fill the absence, eventually. The question, as always, is what conventions will fill it, and whose interests they will serve.

The garment has not yet been tailored. The body — the new medium, with its specific proportions, its specific operations, its specific relationship to the media that preceded it — stands in borrowed clothes that do not fit. The measurements have not been taken. The fabric has not been chosen. The tailor's hands are multiple: the technology companies, the publishers, the regulators, the educators, the practitioners, each cutting cloth according to their own pattern.

When the garment is finished, it will look as though it was always meant to be worn. That is what settled conventions do — they make the constructed look natural, the contingent look inevitable, the negotiated look self-evident. But the garment is being cut now, and the cuts will determine the fit for generations. Gitelman's work is, at its core, a reminder that the fit is never natural. It is always the product of hands, interests, and power.

The medium is unsettled. The conventions are forming. The hands that cut the cloth will determine what the medium becomes.

---

Epilogue

The format is never neutral. That sentence, from a scholar who studies paper and filing cabinets and the institutional archaeology of how information gets organized, has become the phrase I cannot stop applying to everything I build.

I wrote The Orange Pill inside a medium I did not fully understand. I understood the tool — what Claude could do, where it broke, how the conversational interface shaped what I asked for and what I received. What I did not understand was the degree to which the format of the book itself was shaping the story I could tell about the tool. The linear chapter structure imposed a sequence on a process that was not sequential. The single-author attribution implied a creative sovereignty that the collaboration complicated. The conventions of the printed book — conventions so familiar I had stopped seeing them as conventions — were actively shaping what I could say about a mode of production that those conventions were not designed to describe.

Gitelman's framework made this visible. Not as an accusation — she is not in the business of accusations — but as a structural observation. The clothes were borrowed. They fit poorly. And the poor fit was not a failure of the book or the medium. It was the normal, structural condition of every medium in its unsettled period, the period when the conventions that will eventually define the medium are still being invented through the accumulated weight of individual decisions made by people who do not yet know what they are building.

What stays with me is the phrase raw data is an oxymoron — the insistence that no input arrives unmediated, that the training data behind every AI output carries the fingerprints of the institutions that collected it, the languages it was written in, the cultural traditions it represents and the ones it does not. When I described Claude's training data as "the entire history of human thought," I was performing exactly the sleight of hand Gitelman has spent her career exposing. The training data is not the entirety of human thought. It is a subset — shaped by what was digitized, what was available, what survived the filters of platform architecture and copyright law and the economics of web scraping. Calling it everything is the same mystification as calling data raw. It conceals the cooking.

But the finding that unsettles me most is simpler and harder. The conventions that will govern how AI-assisted work is credited, evaluated, priced, and understood are being decided right now — not by philosophers or theorists but by the institutions that have the power to set norms. Technology companies, publishers, copyright offices, academic tenure committees. The decisions these institutions make in the next few years will settle into conventions that persist for decades, and the conventions will feel inevitable once they settle, the way the convention that a book has a single author feels inevitable now — even though it was constructed, through institutional negotiation, during print culture's own unsettled period.

I am one voice in that negotiation. Every practitioner who uses AI tools to build something is one voice. Our accumulated decisions are the raw material from which the conventions will be constructed. But the raw material is not the garment. The garment is cut by the institutions that hold the measuring tape, and the question Gitelman's work forces me to sit with is whether the institutions currently holding the tape are the ones whose measurements I trust.

The format is never neutral. The conventions are never natural. The ground we stand on was always built by someone. The question — the only question that matters during the unsettled period — is who is building it now, and for whom.

— Edo Segal